Oil Ground Up

HiLo Analytics founders discuss AI transforming oil market data analysis and workflow efficiency

1h 0m Listen to the original Episode page
Oil MarketsArtificial IntelligenceData AnalyticsCommodities TradingMarket Data

Summary

Keshav Lohiya and Neil Maheshwari, co-founders of HiLo Analytics, discuss how AI is shifting oil market analysis from manual Excel-based methods to automated, high-frequency data synthesis. They argue that while data volume has exploded, the core challenge remains interpretation and curation, with AI acting as a 'consensus machine' that requires human judgment to generate alpha. The speakers highlight a 'K-shaped' productivity divide where AI users gain massive efficiency gains, and note that market participants are shifting towards high-frequency trading, reducing reliance on traditional supply-demand balances. They also demonstrate the platform's ability to model geopolitical scenarios and test narratives against historical price action.

Episode sections

  1. 0:04 – 4:08

    Introduction and Company Background

    Host Rory Johnston introduces guests Keshav Lohiya and Neil Maheshwari, co-founders of HiLo Analytics. They discuss their backgrounds, with Lohiya having worked at oil trading houses like ConocoPhillips and Ford Motor Company, and Maheshwari coming from Silicon Valley tech. The company was founded in 2019 to address inefficiencies in oil market data analysis.

  2. 4:08 – 9:47

    Evolution of Data and the Shift to AI

    The founders describe the transition from manual Excel-based analysis in 2019 to the current data-rich environment amplified by the pandemic. They emphasize that while data volume has exploded, the core challenge remains data cleaning and interpretation. Maheshwari notes that oil data is smaller in volume but more complex in interpretation than tech data.

  3. 9:47 – 13:08

    Data Curation vs. Synthesis

    The speakers discuss the value proposition of providing raw, curated data versus synthesized answers. They argue that analysts and traders prefer seeing multiple data points and sources to form their own conclusions, rather than receiving a single 'pretty number.' The goal is to make it easier to compare different perspectives, such as varying definitions of Saudi exports.

  4. 13:08 – 17:15

    Market Data Landscape Changes (2014-2026)

    Johnston highlights the shift from relying on lagged IEA/OMR reports and weekly EIA data to high-frequency, granular data like tanker tracking and satellite imagery. Lohiya notes that the 'bull market' in the industry is the availability of data, which causes older data sets to lose value as new ones emerge, creating a dynamic where market drivers constantly shift.

  5. 17:15 – 19:37

    HiLo Analytics AI Platform Features

    Maheshwari details HiLo's AI agent, which provides access to web data, internal datasets, and a Python environment for backtesting and simulations. The platform allows users to ask open-ended questions, run scenarios overnight, and embed the agent into internal systems via API or MCP. Usage is currently high on their platform and WhatsApp bot, with enterprise API integration growing.

  6. 19:37 – 23:51

    Impact on Analyst Workflow and Accessibility

    Lohiya explains that AI automates grunt work, freeing up time for higher-level market thinking and scenario generation. He notes that AI lowers the barrier to entry for complex tasks, such as options risk analysis (gamma/theta), allowing non-quants to perform real-time risk assessments in minutes rather than days.

  7. 23:51 – 27:20

    Use Cases and Data Accessibility Challenges

    Customers use the AI for diverse tasks, including analyzing satellite imagery for refineries and terminals, and maintaining dynamic balances that update with news. Johnston raises the issue of data gatekeeping, noting that organizations like UKMTO are blocking AI crawlers, which increases the value of pre-downloaded, curated datasets.

  8. 27:20 – 32:03

    AI Data Access and Model Diversity

    Speakers discuss how data providers are blocking AI agents from accessing reports, likely due to server load rather than commercial strategy. Neil Maheshwari explains HiLo's multi-model approach, using Anthropic, OpenAI, and Kimi to provide diverse answers and mitigate third-party risk.

  9. 32:03 – 37:05

    Value of Curated Data and AI Training

    Neil argues that high-quality curated datasets are valuable because they serve as the 'gold standard' to train and validate AI algorithms. He suggests that while proprietary data remains valuable, trusted data allows users to build transparent, non-black-box systems.

  10. 37:05 – 42:17

    Market Divergence and Data Proliferation

    Rory Johnston argues that despite more data, market consensus has decreased, with balances differing by millions of barrels. He questions how markets will function if AI provides perfect visibility, suggesting that more data leads to cherry-picking rather than agreement.

  11. 42:17 – 47:06

    Changing Market Participants and Alpha Generation

    Keshav Lohiya notes a shift from physical traders to high-frequency traders, reducing the importance of traditional balances. He states that AI acts as a consensus machine, meaning alpha is generated by deviating from AI consensus, which still requires human judgment.

  12. 47:06 – 51:57

    Testing AI on Geopolitical Scenarios

    Rory tests HiLo's AI on the US-Iran conflict, asking for scenarios. The AI synthesizes news and Polymarket probabilities, outputting a framework of terms, policies, and market impacts, including a 45% probability of a ceasefire within six days.

  13. 51:57 – 53:02

    AI Scenario Testing and Market Reaction

    Speaker_1 tests the HiLo Analytics platform's ability to generate market scenarios, noting it produced reasonable probabilities for 'deal' versus 'muddle through' outcomes without explicit guidance. Speaker_1 observes that the market currently assigns a higher probability to escalation than the model's 15% estimate.

  14. 53:02 – 1:00:07

    Historical Data for Narrative Testing

    Speaker_2 explains that the platform allows users to test narratives against historical price reactions, such as during the Minsk (MOU) agreements. This enables traders to model positions based on whether they believe a current event is a 'head fake' or a repeat of past price action.

Assets discussed

AssetKindStanceContext
Crude oilcommodityneutraloil is still our main business.

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

HiLo Analytics was formerly known as Oilytics and was founded in 2019.

Unidentified speaker (speaker_2) ·1:35

How to verify: Company press releases or SEC filings (if public) or the HiLo Analytics website about page. Verify the founding date and previous name via corporate registry records. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Keshav Lohiya worked at ConocoPhillips and Ford Motor Company before co-founding Oilytics.

Unidentified speaker (UNKNOWN) ·1:55

How to verify: LinkedIn profiles of Keshav Lohiya or professional biographies on the HiLo Analytics website. Cross-reference with public employment records if available. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Neil Maheshwari worked in Silicon Valley, specifically in mobile gaming and e-commerce data infrastructure, before joining HiLo.

Unidentified speaker (speaker_3) ·6:55

How to verify: LinkedIn profiles of Neil Maheshwari or professional biographies. Verify employment history at mobile gaming and e-commerce companies. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

The IEA/OMR report was historically a massive market-moving report but is no longer considered so due to the availability of high-frequency data.

Unidentified speaker (speaker_1) ·13:45

How to verify: Market analysis articles from 2026 vs. 2026 discussing the impact of IEA/OMR releases on oil prices. Compare historical price volatility around OMR releases with recent releases. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Satellite imagery data from Copernicus and NASA was previously difficult to access, requiring terabytes of downloads and manual analysis.

Unidentified speaker (speaker_2) ·25:18

How to verify: Technical documentation or user guides from Copernicus and NASA regarding data access methods. Compare historical data download sizes and processing requirements with current AI-enabled access. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

UKMTO has started blocking AI agent crawlers.

Unidentified speaker (speaker_1) ·26:23

How to verify: Technical logs from AI agents or public statements from UKMTO regarding web scraping policies. Check for changes in robots.txt files or access denial messages from UKMTO websites. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

HiLo Analytics offers an AI agent that can access the web, X, internal data, and run Python backtests.

Unidentified speaker (speaker_3) ·17:44

How to verify: HiLo Analytics product documentation or demo videos. Verify the capabilities of the AI agent by testing the platform or reviewing technical specifications. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

The value of data sets in the oil market is cyclical, with new data causing old data to lose value.

Unidentified speaker (speaker_2) ·15:21

How to verify: Market analysis reports discussing the impact of new data sources (e.g., tanker tracking) on the relevance of traditional data (e.g., IEA balances). Track the correlation between new data releases and market reactions over time. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

AI allows non-quants to perform complex options risk analysis (gamma/theta) in minutes.

Unidentified speaker (speaker_2) ·20:51

How to verify: Case studies or testimonials from HiLo Analytics users demonstrating the use of their AI for options risk analysis. Compare the time taken for such analysis using AI vs. traditional methods. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

HiLo Analytics' platform allows users to update balances in response to news and data releases.

Unidentified speaker (speaker_3) ·18:19

How to verify: Demo of the HiLo Analytics platform showing dynamic balance updates. Verify the functionality by observing how the platform reacts to real-time news or data releases. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Anthropic is more conservative and safe compared to OpenAI, which may provide different answers to the same question.

Unidentified speaker (speaker_2) ·28:20

How to verify: Compare outputs from Anthropic and OpenAI APIs using identical prompts; review model documentation for safety guidelines. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

HiLo Analytics uses a multi-model approach including Anthropic, OpenAI, and Kimi.

Unidentified speaker (speaker_2) ·28:42

How to verify: Review HiLo Analytics product documentation or website for supported AI models. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

UKMTO blocked AI agents from accessing their reports, likely due to server load rather than commercial incentive.

Unidentified speaker (UNKNOWN) ·29:50

How to verify: Attempt to access UKMTO reports via automated scripts; check UKMTO terms of service or technical blogs for updates on API access. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Customs data is considered the 'gold standard' for training AI algorithms on raw AIS and satellite data.

Unidentified speaker (speaker_1) ·32:28

How to verify: Review academic or industry papers on training AI for commodity tracking; check HiLo Analytics methodology whitepapers. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

During the Hormuz conflict, traditional shipping tracking vendors missed volume increases because ships turned off AIS.

Unidentified speaker (speaker_3) ·34:25

How to verify: Review shipping data reports from 2026-2026; check news archives for reports on AIS manipulation in the Strait of Hormuz. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

OPEC secondary source data was offside from implied data regarding UAE production.

Unidentified speaker (UNKNOWN) ·35:09

How to verify: Compare OPEC monthly reports with implied production data from satellite imagery or customs data for UAE. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Oil market balances differed by three million barrels per day before the war started.

Unidentified speaker (speaker_1) ·38:09

How to verify: Review major oil market balance reports (e.g., IEA, OPEC, EIA, private analysts) for the period prior to the conflict. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Open interest in oil futures is declining while open interest in options is increasing.

Unidentified speaker (speaker_2) ·43:06

How to verify: Check CME Group or ICE data for open interest trends in crude oil futures and options contracts. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

HiLo's AI uses Polymarket probabilities to synthesize geopolitical scenarios.

Unidentified speaker (UNKNOWN) ·49:36

How to verify: Test HiLo's AI platform with a geopolitical query; review the output for references to prediction market data. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Polymarket showed a 45% probability of a ceasefire within six days of the podcast recording.

Unidentified speaker (UNKNOWN) ·51:38

How to verify: Check Polymarket historical data for the US-Iran conflict contract at the time of the podcast release. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

The AI platform generated a 15% probability for 'escalation' scenarios without explicit guidance.

Unidentified speaker (UNKNOWN) ·52:29

How to verify: Request a demo or case study from HiLo Analytics showing their scenario generation outputs for recent geopolitical events; compare the model's output against the speaker's stated 15% figure. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

The market was giving a notably higher probability to escalation than 15% at the time of the interview.

Unidentified speaker (speaker_1) ·52:30

How to verify: Analyze implied volatility in oil options markets (e.g., CME Group WTI options) or survey data from energy risk platforms (e.g., S&P Global Platts, Argus) for the period around 2026-11-13 to assess market sentiment on geopolitical escalation. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

The platform allows users to see price reactions during the Minsk (MOU) agreements and when they fell apart.

Unidentified speaker (UNKNOWN) ·53:02

How to verify: Access the HiLo Analytics platform (hiloanalytics.com) to verify if it contains historical price data and event markers for the Minsk agreements (2026-2026) and subsequent market movements. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

AI users will see productivity gains of 100x to 1000x compared to non-users.

Unidentified speaker (speaker_2) ·55:17

How to verify: This is a forward-looking claim; verify by tracking industry reports on AI adoption in commodity trading firms (e.g., from McKinsey, BCG, or specific hedge fund disclosures) over the next 1-2 years. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Software engineers have shifted from typing code to prompting and reviewing AI agent outputs.

Unidentified speaker (speaker_3) ·56:42

How to verify: Review industry surveys on software development trends (e.g., GitHub Copilot usage reports, Stack Overflow Developer Survey) from 2026-2026 to confirm the shift in developer workflows. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify HiLo Analytics' current enterprise client list and API integration growth metrics.
  • Check UKMTO and other data providers' current policies on AI crawler access and traffic gating.
  • Review the performance of AI-generated geopolitical scenario probabilities against actual market outcomes.
  • Compare the productivity claims of '100x-1000x' gains against independent user testimonials or case studies.
  • Assess the reliability of multi-model AI outputs (Anthropic, OpenAI, Kimi) for specific oil market variables.
  • Investigate the shift in oil market participant composition from physical traders to high-frequency traders.
  • Examine the historical accuracy of HiLo's backtesting capabilities using Minsk MOU data as a benchmark.
  • Consider the implications of Jevons paradox on the demand for oil market analysis services.
Full transcriptexpand
0:04
Unidentified speaker (speaker_1): Welcome back to another episode of the Oil Ground Up Podcast. I'm your host, Roy Johnston.
0:10
Unidentified speaker (UNKNOWN): A
0:10
Unidentified speaker (speaker_1): reminder to hit subscribe and leave us a review. Oil Ground Up is distributed in partnership with the Clear Commodity Network at Clearcommodity.net and also the Oil and Gas Global Network, the leading podcast network for oil and gas. Our guests today are Kishav Keshav Lohia. and Neil Maheshwari, co-founders of Hilo Analytics,
0:30
Unidentified speaker (speaker_1): a market data provider, and our discussion focused both on the evolution of the market data business and how the introduction of analytical AI systems has changed the way analysts and the market interact with said data. Keshav and Neil, welcome to the Oil Ground Up Podcast. Thanks so much for joining us today.
0:47
Unidentified speaker (UNKNOWN): Thank you,
0:47
Unidentified speaker (speaker_0): Rory, for having us joining us.
0:49
Unidentified speaker (UNKNOWN): All right,
0:50
Unidentified speaker (speaker_1): so before we get into discussing everything that is going on in the market right now and how your business is trying to tackle it, could you tell us a little bit about the background of Hilo Analytics, or as I initially knew it to be called, Oilytics? You talk about how you guys met, the plan for the business, and how things have developed over
1:10
Unidentified speaker (speaker_1): really the past, was it four or five years? I want to say that oilytics and commodity context. started around similar-ish times, which is how we first met. And I wanted to have you guys on for a conversation forever. And I think on a recent call we were having, just to catch up, some of the topics that you guys were working on really, I think, dovetailed nicely with what we were discussing in the podcast. So I wanted to have you guys on.
1:30
Unidentified speaker (speaker_1): So before we go into all of that, tell us a little bit about yourselves, your backgrounds, and how you got into
1:34
Unidentified speaker (UNKNOWN): this.
1:35
Unidentified speaker (speaker_2): No, thanks, Shore. So yeah, I mean,
1:37
Unidentified speaker (UNKNOWN): very
1:38
Unidentified speaker (speaker_2): similar background. backgrounds. You know, you know, we started a business in 2019. It was called Oiltics Limited, right before COVID started. So, you know, from a timing point of view, couldn't have been a worse timing. But you know, we survived the roughest years first. But, you know, I I I worked for oil trading houses and I worked for, you
1:55
Unidentified speaker (UNKNOWN): know,
1:56
Unidentified speaker (speaker_2): Conical Phillips, Bentford For China. So I've I've been on the other side. And I always had this idea like how analytics was evolving, and I had this idea. And then Neil was working in Silicon Valley at that time in San Francisco. And we always used to speak like what is possible, what's not possible in the tech world.
2:13
Unidentified speaker (UNKNOWN): So
2:13
Unidentified speaker (speaker_2): my domain expertise and Neil's technical skills, you know, we work joined forces.
2:19
Unidentified speaker (UNKNOWN): And
2:20
Unidentified speaker (speaker_2): as a result, now with the way technology has evolved rapidly, our business where it started from to where it's now, you know, it's it's been an absolute journey. Um,
2:29
Unidentified speaker (UNKNOWN): and
2:29
Unidentified speaker (speaker_2): we recently rebranded to high-low analytics, uh, just you know, future-proofing our business, you know, just if we ever move away from oil, oil is still our main business. But um,
2:40
Unidentified speaker (speaker_2): yeah, so that's it's is bit is and we still a small business, you know, you and I, you know, So um
2:46
Unidentified speaker (speaker_2): I think the the industry needs the big players, but also small players. like us, you know, for diversity of opinion, for thoughts, and it keeps the market interesting.
3:08
Unidentified speaker (speaker_1): So tell me a little bit about what the initial idea was
3:13
Unidentified speaker (speaker_1): for Oolytics. OELytics. Like where did you see the comparative advantage? Like for me, my my my initial thought for commodity context was not at all what commodity context began or or became. I'm sure this the story is similar, but I wanted to really fuse some of the kind of more narrative-driven writing that I really enjoyed with some of the very, very
3:33
Unidentified speaker (speaker_1): heavy kind of data tracking, now casting type type work that we saw from other companies at that time. I think two of my inspirations were always RBN energy on the on the on the narrative side and Oil X on the on the data side, when that was before they were acquired by Energy Aspects. So tell me a little bit about how you saw
3:52
Unidentified speaker (speaker_1): Oilytics kind of positioned in that space, and then kind of talk about that reason to,
3:58
Unidentified speaker (UNKNOWN): you
3:58
Unidentified speaker (speaker_1): know, future-proof, future proof as you say, and kind of spread beyond oil. specifically to kind of a broader market data provider, you know, stance.
4:08
Unidentified speaker (speaker_2): Yeah. Um, Joe, so I mean you
4:10
Unidentified speaker (UNKNOWN): know,
4:10
Unidentified speaker (speaker_2): we gotta go back in time. This was I now I think this pre AI world and a post AI world. Okay. So twenty nineteen was pre AI. Yeah you know, um And we're
4:19
Unidentified speaker (speaker_1): gonna get to the AI. So
4:20
Unidentified speaker (UNKNOWN): talk about
4:20
Unidentified speaker (speaker_1): the original idea. idea of
4:21
Unidentified speaker (UNKNOWN): the Just like an idea.
4:23
Unidentified speaker (speaker_2): Like I started my career, you know, my my graduate career in two 2009 thousand nine. for like Conico Phillips. A lot of my first job was to do like seasonality charts in Excel. You know, you did j j Jan to December in one and then the years and used to do seasonality charts for like
4:38
Unidentified speaker (UNKNOWN): trade and it was
4:39
Unidentified speaker (speaker_1): like the vectorized formulas
4:41
Unidentified speaker (UNKNOWN): that you needed. Like
4:42
Unidentified speaker (speaker_1): oh it was
4:42
Unidentified speaker (UNKNOWN): awful. So there
4:43
Unidentified speaker (speaker_2): was a lot of like manual work and and generally to do analysis or organizing data, lots of different companies would were doing it differently. And I had this, you know, this idea that, you know,
4:56
Unidentified speaker (speaker_2): I worked for many multiple different companies. What if we can do this for an industry as a whole and just make it more efficient,
5:02
Unidentified speaker (speaker_0): you know?
5:03
Unidentified speaker (speaker_2): And and and this was before there was a huge data explosion. I think COVID amplified a lot of the data explosion, you know,
5:10
Unidentified speaker (UNKNOWN): if
5:10
Unidentified speaker (speaker_2): like from
5:12
Unidentified speaker (speaker_2): like COVID dashboards. Oh COVID
5:14
Unidentified speaker (UNKNOWN): dashboards
5:15
Unidentified speaker (speaker_1): case data, you know, testing rates, like oh,
5:17
Unidentified speaker (UNKNOWN): I remember
5:17
Unidentified speaker (speaker_1): like
5:17
Unidentified speaker (UNKNOWN): all of the stuff. And real
5:18
Unidentified speaker (speaker_2): time and like real-time flights, all that
5:20
Unidentified speaker (UNKNOWN): stuff,
5:21
Unidentified speaker (speaker_2): you know, 2019, like getting data was just very archaic stuff, you know? And and you know, um
5:27
Unidentified speaker (speaker_2): and oil is like the high frequency stuff was still behind many other industries. So this whole thing was how do we organize and curate different data sets and build it in a nice
5:39
Unidentified speaker (speaker_2): friendly web interface, uh a dashboard business.
5:43
Unidentified speaker (speaker_1): Yeah.
5:43
Unidentified speaker (speaker_2): Um so that was the original idea. And our in behind the scenes where you know Neil and I, uh the data cleaning, the boring, mundane part of the job, but it's still very, very important because end of the day, it's garbage in, garbage out. So if you don't clean your data, the end output is terrible.
6:00
Unidentified speaker (speaker_1): Actually, let me let me actually stop you there. I want to bring Neil in for one. second. Just because again, I haphazardly dabble between the role of analyst and sole developer of commodity context as well. Uh so I kind of do a little I have a foot in both sides and kind of the pitfalls of all. Um so, Neil, when you came, you did not have a background in oil and gas, as I understand. You were kind of more of a
6:20
Unidentified speaker (speaker_1): pure developer, Silicon Valley guy. What was it the first thing that you kind of thought like when you came into this world? of oil data or industry data more broadly? Like, what was your what was your like feel? Like, was
6:32
Unidentified speaker (UNKNOWN): this
6:32
Unidentified speaker (speaker_1): like, oh wow, this stuff is like child's play compared to like the massive data sets of you know open social media or whatever? Or it's like, oh my gosh, this data is like so chunky and so, you know, hard to work with, so messy, I think, as as Kishov was noting. Like, what was your feeling? I always wonder, you know, as
6:49
Unidentified speaker (UNKNOWN): someone
6:49
Unidentified speaker (speaker_1): that knows data and then comes in raw kind of to the oil data space. I'm always curious what their interpretation of
6:54
Unidentified speaker (UNKNOWN): the data space was.
6:55
Unidentified speaker (speaker_3): I was I was working um I started out working in mobile gaming as a game developer and then worked on the data platform side there and then switched over to like an e-commerce company where I started in the data infrastructure side and ended up uh leading like a a small machine learning in and in and data science. team.
7:15
Unidentified speaker (speaker_3): And the type of data you deal with there is very different. It's it's like you mentioned, it's very high volume, it's clicks, uh views, impressions. Um, but that's all
7:28
Unidentified speaker (speaker_3): uh machine generated, you know, kind of consistently formatted. added And and again,
7:34
Unidentified speaker (UNKNOWN): just
7:34
Unidentified speaker (speaker_3): the the scale is like
7:35
Unidentified speaker (speaker_1): Massive but cleaner in many ways.
7:37
Unidentified speaker (UNKNOWN): Massive but
7:37
Unidentified speaker (speaker_3): cleaner. Right.
7:38
Unidentified speaker (UNKNOWN): So
7:38
Unidentified speaker (speaker_3): there's there there is much less uh
7:42
Unidentified speaker (speaker_3): ambiguity actually around each data point you know in in our domain the the data set our la the largest data sets are not are not that large.
7:52
Unidentified speaker (speaker_3): Monthly
7:53
Unidentified speaker (UNKNOWN): data
7:53
Unidentified speaker (speaker_1): is pretty you
7:53
Unidentified speaker (UNKNOWN): know
7:54
Unidentified speaker (speaker_1): even
7:55
Unidentified speaker (UNKNOWN): daily data.
7:57
Unidentified speaker (speaker_3): Even even even tick level prices is actually not that large. If you if you are working on like a small mobile app, your your page views are larger than the entire set of uh exchange exchange prices. You know what I mean? Um however
8:12
Unidentified speaker (speaker_1): it's good it's good to humble us oil analysts on that front. Like, oh no, so much data.
8:16
Unidentified speaker (speaker_3): Yeah. But but however, everything every everything is in the interpretation of of those data points. There's so much that goes into a monthly data point. You know what I mean? Like there there is not that much depth to what one person is clicking on. They clicked, they didn't click. However, every every data point that we deal with has has tons
8:36
Unidentified speaker (speaker_3): of resources put into it. Uh it's it it it's the aggregate of so many people's work and you know there there's a lot of domain knowledge like needed to interpret it. Right. So it it's a it's a very different in my opinion.
8:51
Unidentified speaker (speaker_3): problem. I would say the the the maybe like the core engineering aspects of it are simpler, but but but to use that data is actually you know more involved, I would say.
9:03
Unidentified speaker (UNKNOWN): Yeah,
9:03
Unidentified speaker (speaker_1): because there's like, you know, let's say, like, let's just say as an example, like US gasoline demand or product supplied in March or whatever, right? Like you could have one data from you know, one data set from the EIA, one data set from the IEA, one data
9:19
Unidentified speaker (UNKNOWN): set,
9:19
Unidentified speaker (speaker_1): then that those aren't always, you know, US is actually pretty close, but like for
9:22
Unidentified speaker (UNKNOWN): like
9:22
Unidentified speaker (speaker_1): most of the countries, you could have four different observations for the same data point.
9:28
Unidentified speaker (UNKNOWN): You
9:28
Unidentified speaker (speaker_1): could have slightly differing definitions between those, between those surveys or those methodologies. You have revisions. So vintage
9:35
Unidentified speaker (UNKNOWN): of data
9:36
Unidentified speaker (speaker_1): matters a lot more than in like say user-generated. data. Like, so how have you guys like in where do you see the value at? Because I think I always struggle with this as well. Like, do you want to provide
9:47
Unidentified speaker (UNKNOWN): and
9:47
Unidentified speaker (speaker_1): like again, and this comes out like what is who is the end user of the data? Is it analysts? Is it decision makers? Is it traders? Is it whomever? There's so much different use there. Like an analyst wants all the raw input, right? They want to do the analytics themselves, they want
10:01
Unidentified speaker (UNKNOWN): to
10:02
Unidentified speaker (speaker_1): see the sausage. is made. Traders sometimes do. Sometimes they just want
10:06
Unidentified speaker (UNKNOWN): the
10:06
Unidentified speaker (speaker_1): answer or whatever, just to make you know simpler.
10:08
Unidentified speaker (UNKNOWN): So
10:08
Unidentified speaker (speaker_1): like where do you guys see it between like curating differences in this data versus like synthesizing, you know, purer, you know, cleaner data points for use?
10:19
Unidentified speaker (speaker_3): I I mean I
10:19
Unidentified speaker (UNKNOWN): think
10:20
Unidentified speaker (speaker_1): either of
10:20
Unidentified speaker (UNKNOWN): you,
10:20
Unidentified speaker (speaker_1): I think is it it would be interesting,
10:21
Unidentified speaker (UNKNOWN): or actually both.
10:22
Unidentified speaker (speaker_3): I guess I'll definitely start. Like I I think that's an e I like as Kishan mentioned before, yeah. That's that's that's probably something that's evolving tremendously. Um we do really we resonate very well with people who want to uh dig in and they want to see several
10:42
Unidentified speaker (speaker_3): different numbers and come up with their own conclusion, you know. We do really well with with with with that persona person, whether they're a trader or an analyst, to be fair. Um I think we I think people who really question what they're seeing, who want to come up with their own conclusions, who don't want one pretty number.
11:02
Unidentified speaker (speaker_3): Um
11:02
Unidentified speaker (UNKNOWN): it's
11:03
Unidentified speaker (speaker_3): it's kind of hard
11:03
Unidentified speaker (UNKNOWN): to do. It's so comforting.
11:04
Unidentified speaker (speaker_1): Pretty numbers are so
11:05
Unidentified speaker (UNKNOWN): comforting. Right.
11:06
Unidentified speaker (speaker_3): And it's like a little irony as we do we do curate data, but but the but the the goal to the goal of data curation is not to smooth the complexity, right, of of of what you're looking at, right? It is to just make it possible for you to see those n different numbers or those n different perspectives easier, right?
11:22
Unidentified speaker (speaker_2): Yeah. Yeah. And like we started off with like our business our dashboard business in the beginning was like how do we simplify
11:28
Unidentified speaker (UNKNOWN): or
11:29
Unidentified speaker (speaker_2): curate the data set of things. You look at a huge Jodi database. You
11:32
Unidentified speaker (UNKNOWN): know,
11:33
Unidentified speaker (speaker_2): this was like 200 odd gigabytes, everyone's downloading the monthly file, everyone's doing the same thing every month. Um, we started off as that, but now what we found out is that one data set on its own is not that important. Is when you combine one data set with along 50 other data sets, you
11:48
Unidentified speaker (UNKNOWN): know,
11:48
Unidentified speaker (speaker_2): this could be Chinese customs data, this could be
11:52
Unidentified speaker (speaker_2): daily flight data, um, and and combining these data sets and when and if you want it, that's the power. You know, that's the that's that's when it gets useful. And also like different vendors as well. You know, sometimes they just give a black box. Uh, but now with with AI or with different data sets, you can deconstruct any kind of input or
12:12
Unidentified speaker (speaker_2): uh that goes into it, you know. Um
12:17
Unidentified speaker (speaker_2): and like with even like balances, you know, you don't you don't wanna you don't wanna absorb someone else's balances without questioning it because there are so many assumptions, so many different black boxes, and this is when like so this is, you know, we have we have
12:28
Unidentified speaker (UNKNOWN): you
12:28
Unidentified speaker (speaker_2): know clients from like traders or analysts or whatever. And it depends how you want to use the data. You know, some people just want like a a nice pretty dashboard. Or some people want to just drill down and and and deep dive on any numbers or even compare different data sources where those one number is can have three different numbers, you know, from the
12:48
Unidentified speaker (speaker_2): same Saudi exports, for example, can have three different numbers for the same month.
12:53
Unidentified speaker (UNKNOWN): Uh Right.
12:53
Unidentified speaker (speaker_1): And I think like one of the things that's interesting is like how how vastly this industry has changed, even in my short time in it. Like I got my first real job in this space at Scotiabank in 2014. Um at that stage, in 2014, this was before many of the tanker tracking companies had emerged,
13:13
Unidentified speaker (speaker_1): certainly before people were using or at least it was commercially available, like satellite derived inventory estimates or whatever.
13:21
Unidentified speaker (speaker_1): Kind of the industry still lived and breathed by weekly US, you
13:25
Unidentified speaker (UNKNOWN): know,
13:25
Unidentified speaker (speaker_1): weekly petroleum status report data from the EIA. Uh, and then more generally on like two month lagged IEA balances, largely, but people would do comparisons or whatever. But largely it was the IEA IAOMR OMR on a two-month two month lagged basis. And that was kind of how the entire industry ran. Like the IAOMR was a massive market-moving report.
13:45
Unidentified speaker (speaker_1): It is no longer a massive market-moving report. And I think that even before the AI stuff, we evolved so dramatically, both in terms of the data available, but also I think the skill sets and um kind of techniques available to analysts to use.
14:03
Unidentified speaker (speaker_1): much broader data. So like like I even before AI, I was built, I was trying to build my own balances to reconstruct some of the stuff from the ground up. And I don't think I would have been able to do that. I couldn't have done that back in 2014. Um, at least not the skills I had then.
14:16
Unidentified speaker (UNKNOWN): And
14:17
Unidentified speaker (speaker_1): I think that's one of the interesting things. So like even again, kind of final question before we get to the AI is like, how did you even see that period evolving? Because even like. the data space, the backdrop, the context, if you will, in which we were all operating, it became much more high frequency, much more higher fidelity.
14:37
Unidentified speaker (speaker_1): Like people weren't really talking about global balances in the same way. It was regional, it was product level, it was whatever. There's a lot more granularity. But like, how did that kind of how did that evolution affect how you guys look at the market and how your clients looked at the market? And also like how all of this just generally affected price formation? Because again, I OMR was no
14:57
Unidentified speaker (speaker_1): longer the market moving report. Even EIA weeklies weren't really that market moving anymore. Whereas like there used to be a lot of alpha and like,
15:04
Unidentified speaker (UNKNOWN): you
15:04
Unidentified speaker (speaker_1): know, predicting what the EIA print was going to be. Now it's kind of like, eh, whatever.
15:10
Unidentified speaker (speaker_2): Yeah. The the The only real bull market in our industry is the ability availability of data. It's not flat price, you know.
15:21
Unidentified speaker (speaker_2): Um yeah, I mean the sheer amount of data that's come out, it's it's crazy. It's it's exponential, you know. It's um and also it's a chicken and the egg game. You know,
15:29
Unidentified speaker (UNKNOWN): uh
15:30
Unidentified speaker (speaker_2): as the market absorbs more and more data, the existing data sets lose its value. So for example, right now, uh with so much of the market being moved by headlines, uh, there's a lot of analysis being done on sentiment analysis or X or Twitter, um, COT positioning, you know, which which back then
15:50
Unidentified speaker (speaker_2): was just one of the data sets. So the market's constantly evolving and it imp it it puts different weightings on different data sets at different times. And no one knows that weighting waiting at real time, you know, it could be ship tracking, it could be positioning. And that's what makes people want to consume more and more data. It's just you don't know when that
16:09
Unidentified speaker (speaker_2): waiting switches on on the various data sets. Um
16:15
Unidentified speaker (UNKNOWN): yeah,
16:15
Unidentified speaker (speaker_2): so like fundament like you
16:16
Unidentified speaker (UNKNOWN): know,
16:17
Unidentified speaker (speaker_2): a lot of people are fundamental, like you know, I think you and I, Aurora, like fundamental like analysts, analyst, you know, which was which
16:24
Unidentified speaker (UNKNOWN): we used to hold,
16:25
Unidentified speaker (speaker_2): which we used to hold with a proud. you know, with a proud badge. Uh for rightly or wrongly reasons. You know, fundamental is just one aspect of the white oil markets, and we have to cater to that other growing data sets.
16:41
Unidentified speaker (speaker_2): Yeah.
16:41
Unidentified speaker (UNKNOWN): All
16:42
Unidentified speaker (speaker_1): right. So let's let's get to this, you know, again, the the topic of our phone call that that that really presaged this this uh podcast conversation, which is over the past year or two, uh Hilo Analytics has been aggressively pushing down more an AI-enabled, agentic
17:02
Unidentified speaker (speaker_1): analytical platform. So let's talk, let's before we kind of get into what it means, like tell me what you guys think it means. Like what is what is the pitch to clients now and how has that evolved since rolling this piece out?
17:15
Unidentified speaker (UNKNOWN): Yeah,
17:15
Unidentified speaker (speaker_3): absolutely. So I
17:16
Unidentified speaker (UNKNOWN): mean
17:16
Unidentified speaker (speaker_3): it's it depends who who we're we're we offer uh a few different solutions depending on who we're talking to. Um at It Simplest, we offer
17:24
Unidentified speaker (UNKNOWN): you
17:24
Unidentified speaker (speaker_3): know a chat interface where our clients can ask any questions and get get back answers. Now, that's kind of what it looks like on the surface. What's r what what it really gives you access to access to is a mini quant, a mini kind of like analyst and in a box. So our agent has full access to the web, X, uh,
17:44
Unidentified speaker (speaker_3): all our data sources, a variety of other uh uh, you know, whatever it can find online, as well as like a s a small uh Python environment, um which it can run back tests, it can do simulations, uh and and
18:00
Unidentified speaker (speaker_3): that's that's super useful, you know, for in individuals, uh traders analysts who just need to get answers. And and and what well, that was the beginning where people were kind of asking simple questions. What we see now is people doing real work. You know, like I an example I love is we have a some customers who have balances that they now have their
18:19
Unidentified speaker (speaker_3): agent updating for them in response to news, data releases, uh uh scenarios as well, if they kind of want to play. with that. Um, we also offer uh tools for you know enterprises to embed our agent, embed our data uh into their systems. So, you know, they may have an internal
18:39
Unidentified speaker (speaker_3): deployment and they can access us via uh an API or an MCP, which is the AI uh version of of an API. Um and and that kind of just like supercharges their their their application. Uh lastly, like we've we've made this available wherever they work. So if they want to talk to it via WhatsApp,
18:59
Unidentified speaker (speaker_3): if they want to talk to it, we recently added Zoom chat, which is a new chat. I didn't realize
19:04
Unidentified speaker (UNKNOWN): we had
19:05
Unidentified speaker (speaker_3): a chat. Email. So you know it feels more and more and more comfortable. But I think the thing I really want to emphasize is that in the beginning when we started this a year and a half ago, people were asking simple questions. It's almost like information retrieval. You know, what are Chinese refinery runs? You get back an answer. Now, what we're talking about is asking questions
19:25
Unidentified speaker (speaker_3): that are open-ended, scenario based, exploratory. And then when they have something they like, they can delegate it and that that agent can run that in the background for them, you know, overnight or or or send it to whatever event.
19:37
Unidentified speaker (speaker_2): And and this goes back
19:38
Unidentified speaker (UNKNOWN): to
19:38
Unidentified speaker (speaker_2): like when I started the industry in 2009, there's so much of my daily hours was spent on grunt work. You know? And now once you can like delegate that or offload that, you can spend so much more time on like thinking about the markets or or analyzing, you know, the AI is very good at
19:58
Unidentified speaker (speaker_2): doing grunt grant work, but it also frees up your time. So where now with AI, your human judgment becomes more and more important. So imagine all those hours, like, you know, when you're when you join the industry, or even now, like, you know, doing presentations, different things. If you can
20:13
Unidentified speaker (speaker_2): get in automate, not just automate stuff, but it you're you're it's not just getting your hours back, but you're also coming up with scenarios and different new lines of work, which was previously unthinkable. And even like uh, for example, like before. we started off as a dashboard business, you know, now with front end, everything is so easily accessible, it doesn't matter if
20:33
Unidentified speaker (speaker_2): you're at your your your your desktop or everything is information on the go now. Um
20:42
Unidentified speaker (speaker_2): and and and also like having a as Neil said, uh, you know, like I I use I love options, but I never got into the details of like gamma
20:50
Unidentified speaker (UNKNOWN): or
20:50
Unidentified speaker (speaker_2): theta.
20:51
Unidentified speaker (speaker_2): because that was like a like a bit too um in depth for me. But now you can have a quant at your fingertips. You can do real time uh risk analysis or you know, uh how would my risk increase if I put this position on? And you're getting such real time feedback uh in days, up in minutes, which you would have taken you
21:11
Unidentified speaker (speaker_2): days before. Um I mean
21:13
Unidentified speaker (speaker_3): I think
21:13
Unidentified speaker (UNKNOWN): I think in in some
21:14
Unidentified speaker (speaker_3): of your your quant examples, I think that's work you just would have never considered, right? I mean, there's there was no way you were ever going to teach yourself enough uh
21:23
Unidentified speaker (speaker_3): Python and options theory to get an answer, right?
21:30
Unidentified speaker (speaker_1): And how and how would you say that this has, you know, when you're speaking with your clients, like how are they using this data? I know you kind of gave a couple of examples, but like in in your discussions with them, like what
21:42
Unidentified speaker (UNKNOWN): and
21:42
Unidentified speaker (speaker_1): I am also curious, like how many of them are doing this like from your chat window versus plugging the you know um you know the the APIs directly into their existing systems? Like, how does it really work with the normal workflows here? Because I'm I'm it's like I use AI systems in my
21:57
Unidentified speaker (UNKNOWN): work,
21:58
Unidentified speaker (speaker_1): um, but I kind of am honestly constantly changing how I use them. I think I have not yet settled on a f on a flow. Uh and I'm very curious like to see how others have done so.
22:08
Unidentified speaker (speaker_3): Yeah. So I guess Roy, just to to back up, you are, you know, if if there's a range of how
22:16
Unidentified speaker (UNKNOWN): uh
22:17
Unidentified speaker (speaker_3): gung ho people are about like embracing a new technology, you're pretty far to one edge, right? You, you know, since five, six years ago have been building your own systems. And so you're very comfortable with that, right? So, like, I think like just you are probably in the population, you know, pretty far to one side. Um, even though you saw yourself short, right?
22:37
Unidentified speaker (speaker_3): Um our our customers, uh just in terms of how they use it, you know, when we talk to people, there's often a lot of initial
22:46
Unidentified speaker (UNKNOWN): the
22:46
Unidentified speaker (speaker_3): people that we're talking to are the AI champions in their organization. They are they're they're the people. like you, right? Because they're they're the ones who are most eager, of course, right? We're we're we're we're gonna speak with them first. And so they'll have a lot of enthusiasm for bringing us on and using us in their internal systems and using, you know, our platform because you know, they're they're they're
23:06
Unidentified speaker (speaker_3): they're just curious and they're hungry. And right now we see whether they have access to both. We still see the majority of the use in our platform. And because there is there's a still a pr there's a pretty large gap between getting a working internal system and and kind of getting what what we've what we've put together, you know? Um
23:26
Unidentified speaker (speaker_3): and I think a lot of that polish sort of builds up. I I I don't I anticipate in the future, as companies' internal systems get a little bit more advanced, they'll they'll take better advantage of of of of of sort of plugging our our building blocks in. Um but right now I think most people are yeah in our platform or on our WhatsApp bot, I would
23:46
Unidentified speaker (speaker_3): say. And then the use cases are are I I think what I find so interesting is how
23:51
Unidentified speaker (speaker_3): broad the use cases are. This is really a you know a mirror to our customers' creativity in in in in some sense, right? Unlike before, where we were delivering sort of like a well-defined tool. the the more curious our customers are, the the better questions they ask, the the better results they get. So you know we have people um
24:12
Unidentified speaker (speaker_3): who are doing fascinating things with satellite imagery where, you know, every time there's a new s satellite pass, they they they are, you know, watching lots of of of points of of interest some of these are refineries some of these are terminals different berths
24:32
Unidentified speaker (speaker_3): um a a another use case i really like is i really like the the we have a customer who has has uploaded some balance and now that balance updates daily in response to the news or data releases right I think I find that so interesting um
24:50
Unidentified speaker (UNKNOWN): Kish
24:50
Unidentified speaker (speaker_3): have mentioned and I and I and I think that the back testing examples are really interesting.
24:54
Unidentified speaker (UNKNOWN): You
24:54
Unidentified speaker (speaker_3): know, you have you have
24:55
Unidentified speaker (UNKNOWN): an
24:55
Unidentified speaker (speaker_3): idea for a trading strategy. Uh oh sorry go for it, Kish.
24:58
Unidentified speaker (UNKNOWN): No,
24:58
Unidentified speaker (speaker_2): sorry, I was saying and on that satellite imagery, Rory, like back in the day, that satellite imagery data set would have been a six figure plus uh subscription. Um so you know inflation is everywhere, but one thing we are getting disinflation could be on data costs because, you know we as you said from like 2014 onwards, a lot of the data sets were
25:18
Unidentified speaker (speaker_2): guarded. Like even getting satellite imagery, you know, there's a lot of satellite imagery from Copernicus or NASA, but retrieving that pre AI, you would need to download terabytes
25:29
Unidentified speaker (UNKNOWN): and
25:29
Unidentified speaker (speaker_2): have like an army of people to analyze this stuff. Now you can get satellite silent imagery on demand. Um you just gotta got put to a location and it will it will retrieve the correct lat long and it will also give you different time lapses of images and it will analyze the change. So just in that example,
25:49
Unidentified speaker (speaker_2): even though the information was there, Copernicus has been there for many years, but now the ability to analyze and retrieve it when and if you want is practically free and achievable. So
26:02
Unidentified speaker (UNKNOWN): Yeah,
26:03
Unidentified speaker (speaker_1): that's
26:03
Unidentified speaker (UNKNOWN): because
26:03
Unidentified speaker (speaker_1): one other thing I was curious about. Like one thing that again, like so one of the things that I had been using chatbots or whatever for was collecting parsing data on and reports on shipping attacks and Hormuz. And one of the things I noticed very recently was, I want to say like over the last past couple weeks, was that
26:23
Unidentified speaker (speaker_1): UKMTO has started blocking the these agent crawlers. Um, so one of the things that I'm like, I've I'm always someone that even before that, I'm not, I don't really have it just do stuff. I tend to curate the stuff it's looking at and then have it do the you know, have it do it on a cleaner data set.
26:40
Unidentified speaker (UNKNOWN): But
26:40
Unidentified speaker (speaker_1): I always already download these files. But when when this was happening, I was like, oh wow, that's interesting. Now I there's more value to the fact that it I already downloaded this stuff and had it organized and sorted. Like, how much do you see that changing? Because I think one of the things that people keep talking about is that like the era of the open internet is coming to an end, that the kind of proliferation
27:00
Unidentified speaker (speaker_1): of agents, the kind of needing to gate and guard things in a way that you never had to before, uh, because just the the the cost of overuse, the cost of access is now so, so, so, so low. Um, that places like UKMTO, which historically would have never had protections
27:20
Unidentified speaker (speaker_1): against their sites, now all do. And more and more websites, I find, are putting up some kind of gate. Um, like, how do you see that changing the value prop of like the exhuriated data you guys have? Like, one of the things I think is interesting is that that becomes more valuable as general agents have less and less ability to
27:40
Unidentified speaker (speaker_1): access that data open source.
27:42
Unidentified speaker (speaker_3): Can I ask you one before before I answer? Can I see one question? Why do you think they have decided to block agents from accessing their reports, right? As a right, that there's no they don't have a commercial incentive. Is it is it a
27:56
Unidentified speaker (UNKNOWN): lot of people?
27:56
Unidentified speaker (speaker_1): It's true.
27:56
Unidentified speaker (UNKNOWN): It's true. That's
27:57
Unidentified speaker (speaker_1): And I actually don't know the reason. All I know is that
27:59
Unidentified speaker (UNKNOWN): is that, you
27:59
Unidentified speaker (speaker_1): know, all you know, a couple weeks ago, Chat ChatGPT GPT was all of a sudden started saying, I actually can't access that. If you download
28:05
Unidentified speaker (UNKNOWN): the
28:05
Unidentified speaker (speaker_1): files, I can do that for you, but you know, I can't access it. And that was interesting to me.
28:12
Unidentified speaker (speaker_1): Yeah, you're saying shot.
28:13
Unidentified speaker (speaker_2): Neil, can
28:13
Unidentified speaker (UNKNOWN): I give a can I give a non
28:14
Unidentified speaker (speaker_2): technical answer before you give a technical one?
28:16
Unidentified speaker (UNKNOWN): Yeah,
28:16
Unidentified speaker (speaker_3): yeah. I don't I'm I'm asking like I don't
28:17
Unidentified speaker (UNKNOWN): have an answer.
28:18
Unidentified speaker (speaker_3): I'm just saying there's no
28:18
Unidentified speaker (UNKNOWN): commercial
28:19
Unidentified speaker (speaker_3): reason why they wouldn't.
28:20
Unidentified speaker (speaker_2): So I'll tell you our building blocks, right? It means like
28:22
Unidentified speaker (UNKNOWN): the
28:22
Unidentified speaker (speaker_2): brain behind it, but our building blocks saying, like, you know, we fed our collected our data sets to. anthropic. So in the beginning, so anthropic, you know, every AI is different. So anthropic is more on the conservative side. So if you ask a certain question, like just like what OpenAI did for you, uh, it could give you a very
28:42
Unidentified speaker (speaker_2): conservative or safe answer. So OpenAI in that example didn't go and and and go through the UK MTO. What we've now done, we've done a multi-model. So the users can connect, or they can change in our web chat anthropic, Kimi, which is Chinese version, uh, or OpenAI, or different, and you know, this is this is a huge OpenAI race and and the anthropic lead is
29:02
Unidentified speaker (speaker_2): is converging. But my point is that if you ask the same question to the Chinese version, the Kimi, you would get a much different answer. Because
29:08
Unidentified speaker (UNKNOWN): they I
29:08
Unidentified speaker (speaker_1): I'm sure, I'm sure it's not quite as respectful of uh of stated guardrails.
29:13
Unidentified speaker (speaker_2): So anyway, I I so so my point is that you know, even when you're running a business, you know, we don't want to be at the whims of just one AI model, because then you're you have a
29:22
Unidentified speaker (speaker_2): a third-party risk. So so uh my point is we we give multimodal options uh not just from like a better answer, but also a diversity of answers, just like we our original point was the market wants a diversity of answers, not just from better answers. vendors, consultants, but also AI models. But I'll let Neil
29:42
Unidentified speaker (speaker_2): give a more technical answer why you've got that issue.
29:45
Unidentified speaker (UNKNOWN): Well, and I
29:46
Unidentified speaker (speaker_1): I I I mean on the technical side, I just assume this is a like a traffic gating thing. That
29:50
Unidentified speaker (UNKNOWN): they just don't
29:51
Unidentified speaker (speaker_1): want millions
29:52
Unidentified speaker (UNKNOWN): of agents
29:52
Unidentified speaker (speaker_1): kind of ripping down PDFs willy-nilly and kind of gumming up their system. Which I think is completely understandable. Um but it is tricky because
30:00
Unidentified speaker (UNKNOWN): then it'cause it
30:01
Unidentified speaker (speaker_1): and I think one of these things, it then it it you know, to the you know, OpenAI versus Kimmy kind of uh example, like it it also then tends to almost create incentives for bad behavior. Like like it like it disincentivizes people acting, you know, cordially. Uh I mean even before this, like there was always those what was it,
30:21
Unidentified speaker (speaker_1): uh Robot robots robots.txt dot txt or whatever, Neil,
30:24
Unidentified speaker (UNKNOWN): like where like you would have your kind of
30:25
Unidentified speaker (speaker_1): stated H. API scraping. And it was kind
30:28
Unidentified speaker (UNKNOWN): of just
30:28
Unidentified speaker (speaker_1): like this unspoken like don't
30:30
Unidentified speaker (UNKNOWN): be a
30:30
Unidentified speaker (speaker_1): jerk and like follow what people want.
30:33
Unidentified speaker (speaker_3): Right.
30:34
Unidentified speaker (speaker_2): So I I guess I guess
30:36
Unidentified speaker (speaker_3): my feeling is that like these are probably growing pains, right?
30:44
Unidentified speaker (speaker_3): Because you UK MTO doesn't have a commercial incentive to hide the information. These are probably growing. growing pains from the rise of agentic traffic. And they are likely there's a good chance that they're just they just have some firewall who rolled out an update. You know what I mean? It might be that as benign as that, right? Now I
31:04
Unidentified speaker (speaker_3): guess
31:04
Unidentified speaker (UNKNOWN): I think on the on the
31:04
Unidentified speaker (speaker_3): commercial
31:05
Unidentified speaker (speaker_1): Cloudflare update or
31:06
Unidentified speaker (UNKNOWN): whatever. Right,
31:07
Unidentified speaker (speaker_3): exactly. Right. So I don't I
31:08
Unidentified speaker (UNKNOWN): don't know
31:08
Unidentified speaker (speaker_3): enough to to answer, but I'm I'm I feel some of these things are growing pains. You know, some of these things are
31:13
Unidentified speaker (UNKNOWN): uh
31:13
Unidentified speaker (speaker_3): because because it's like you mentioned uh anyone who really wants to get it can get it, right? Can get
31:18
Unidentified speaker (UNKNOWN): can get it. For
31:19
Unidentified speaker (speaker_1): sure, right? I think
31:20
Unidentified speaker (UNKNOWN): they haven't they
31:20
Unidentified speaker (speaker_1): haven't but I think even even in that case, like I mean, it's far less taxing on UKMTO servers
31:26
Unidentified speaker (UNKNOWN): for
31:26
Unidentified speaker (speaker_1): me to download a couple PDFs and then you know have
31:29
Unidentified speaker (UNKNOWN): them look at it. Right,
31:30
Unidentified speaker (speaker_3): exactly.
31:30
Unidentified speaker (UNKNOWN): Versus the
31:31
Unidentified speaker (speaker_1): same thing querying that same page like a hundred times
31:33
Unidentified speaker (UNKNOWN): in a minute.
31:34
Unidentified speaker (speaker_3): Yeah, yeah.
31:34
Unidentified speaker (UNKNOWN): So
31:35
Unidentified speaker (speaker_3): there's there's tech solutions on both sides. Eventually they'll cache cash it. and it won't cost their servers. And in two, three years' time, like that won't be issued. But I think on the commercial side, you know, they've or or on on the proprietary side, that's a very good question because
31:50
Unidentified speaker (speaker_3): it is incredibly easy to reverse engineer many things with with with AI tools. User interfaces are are trivial at this point. Um
32:03
Unidentified speaker (speaker_3): You know, this this
32:05
Unidentified speaker (speaker_3): kind of tying into the other part of your question, one of the values of
32:10
Unidentified speaker (speaker_3): good curated data sets is that they can be used to bootstrap systems. So for example, if you were a large company and you had a good
32:21
Unidentified speaker (speaker_3): several years of of customs data and you had a bunch of raw AI. AIS signals.
32:28
Unidentified speaker (speaker_1): Yep.
32:29
Unidentified speaker (speaker_3): And maybe some bank of satellite imagery. You could today take an agent and say,
32:37
Unidentified speaker (UNKNOWN): these
32:38
Unidentified speaker (speaker_3): cust this customs data is the gold standard. It is correct. Design me a few algorithms that take my raw AIS data in satellite imagery and output the same monthly aggregates.
32:50
Unidentified speaker (speaker_1): Yeah.
32:51
Unidentified speaker (speaker_3): Right. And,
32:53
Unidentified speaker (UNKNOWN): you
32:53
Unidentified speaker (speaker_3): know, you might spend a couple hundred, a couple thousand bucks, probably need a few developers to guide this process, but you will end up with an algorithm that works fairly well. You know? So
33:08
Unidentified speaker (speaker_3): as people who provide um
33:12
Unidentified speaker (speaker_3): as people who provide these very high quality data sets, that there there's there's additional value to those, I think, in in the AI land, in the AI sorry, you know, world. Um
33:26
Unidentified speaker (speaker_3): I think there will still always be value for for the expensive proprietary data sets as well. But customers can be far more aggressive in terms of what they try to tweak or do in-house. I think the sort of like that's like the negative framing of that. The positive framing of it maybe is more that like
33:44
Unidentified speaker (speaker_3): with good data sets that you can trust, you can make fewer of your systems black boxes, right? So you can take the
33:52
Unidentified speaker (UNKNOWN): the
33:53
Unidentified speaker (speaker_3): the vendor's point of view, you can take the various building blocks that you have and you can take the uh data that you trust the most, maybe if it's if it's the National Statistics Org or whoever. and you can say, you know, let me build a view that incorporates all three, basically. Right.
34:09
Unidentified speaker (speaker_2): And this becomes very important where the error rate is getting larger in the oil markets. So even with more transparency, more technology, more tracking, we are still in a world where the tracking rate or the error rate is just getting higher and higher.
34:25
Unidentified speaker (speaker_3): Yeah.
34:25
Unidentified speaker (UNKNOWN): So
34:25
Unidentified speaker (speaker_2): it's the it's the ultimate category. cat and mouse game. Like even with this, even this Hormoose conflict, the the traditional shipping ship tracking vendors missed the run-up with in volumes uh picking up because they switched the AIS off and satellite imagery became more important. So we
34:42
Unidentified speaker (speaker_2): is the ultimate cat and mouse game, no matter how smart or how sophisticated you are, there will be a time when you will miss a big chunk of the market. I
34:50
Unidentified speaker (speaker_1): think even like I always think, I mean, even just even before all of this went down, uh, we were, you know, one thing everyone was following was OPEC quotas, production, what was going on with the UAE in particular. And I think one of those big questions was like, I think like all of the implied data showed much higher,
35:09
Unidentified speaker (UNKNOWN): you
35:10
Unidentified speaker (speaker_1): know, Emirate production. And the official OPEC secondary source data that for my entire career was kind of treated as the gold standard for OPEC tracking was all of a sudden very clearly offside from these other data sources. So I think it's very interesting to kind of say like, I always like how some of the AI agents
35:29
Unidentified speaker (speaker_1): they kind of coily point to this stuff. They're like, well, that's It's funny. Or, like,
35:35
Unidentified speaker (UNKNOWN): you
35:35
Unidentified speaker (speaker_1): know, oh, you should probably reconcile that or figure out which way. one you trust. But I think this kind of brings it to this other point, which is
35:43
Unidentified speaker (UNKNOWN): I
35:43
Unidentified speaker (speaker_1): think the big question here. Um, and again, as a as you as you notice, Kishaba, uh, uh I'm a bit of a relic of the of the fundamental analytics space, um, humans overall, rah-rah. But
35:58
Unidentified speaker (UNKNOWN): in
35:59
Unidentified speaker (speaker_1): a world, let's say, again, I we I grew up like a huge trekie, Treky, right? Like, so like the idea of a super intelligent crew member in like data wasn't like that was always like something I hoped for. Like I would say, like, this is like we're getting towards this kind of world where you can ask a computer, like the computer on Star Trek, you can just say, computer, what was this fact? And
36:19
Unidentified speaker (speaker_1): it will just tell you the fact. It knows everything. We are kind of getting to that kind of stage, but obviously,
36:25
Unidentified speaker (UNKNOWN): there
36:26
Unidentified speaker (speaker_1): were still like the computer didn't make all the decisions, data didn't make all the decisions. And I'm curious to see, like, even in a world where I'm still very frankly skeptical on some of the
36:38
Unidentified speaker (speaker_1): statements of competence of some of these systems in all, you know, in all weather, but no doubt they are getting truly extraordinarily impressive. Let's fast forward, say even three, four years.
36:52
Unidentified speaker (speaker_1): and assume that anywhere near this kind of exponential capacity kind of continues to develop, puts us in like a completely different world then, where these systems do have like almost perfect knowledge of what can be known.
37:05
Unidentified speaker (speaker_0): Yeah.
37:06
Unidentified speaker (speaker_1): How does that change how
37:10
Unidentified speaker (speaker_1): markets work? I think like let's let's look at oil specifically. The crude oil market. Let's say we had perfect visibility on
37:19
Unidentified speaker (speaker_1): any piece of it. Like we really don't have a perfect visibility on any piece of it right now. Let's assume we had perfect visibility on any one piece of it. How does that change things and where does the diversity begin to come from? Like I would argue, like right now, in some ways,
37:31
Unidentified speaker (UNKNOWN): one
37:32
Unidentified speaker (speaker_1): thing I've always noted with like the proliferation of this alternative data is that on one level, you would have assumed that. everyone would all of a sudden start having more and more closer and closer balances, that you would begin to develop a deeper consensus.
37:49
Unidentified speaker (UNKNOWN): And
37:49
Unidentified speaker (speaker_1): in my humble opinion, the opposite has happened: that you've pro you've created way more opportunities for cherry picking. That when you have when you had a hundred data points, everyone kind of had to get to somewhere near the same level. But when you have a million data points, there are so many opportunities to pick the most bullish interpretation of every single decision gate, or the most bearish, or whatever, right?
38:09
Unidentified speaker (speaker_1): Which is why I think like the level of debate and the and even the like even going into this, people were like three million barrels a day off of each other's balances for this year before the war started. Like that, if you were just to say, like, we're looking from the outside, that shouldn't be possible with given the level of data fidelity. at least we think we have right now. So
38:29
Unidentified speaker (speaker_1): like talk to me
38:31
Unidentified speaker (speaker_1): about how you think that changes and how you think even this conversation would sound different in three or four years. Um I'm just very, very curious because I have no idea.
38:42
Unidentified speaker (UNKNOWN): I
38:42
Unidentified speaker (speaker_3): mean, I have
38:43
Unidentified speaker (UNKNOWN): one, I have
38:43
Unidentified speaker (speaker_3): I have one I have a few I have two two thoughts. Yeah. Like my my first thought is that there are just still there are well, while well, well. while you said there's these large divergences on people's balances, there are it is still
38:56
Unidentified speaker (speaker_3): uh there are still not enough people
39:00
Unidentified speaker (speaker_3): doing the work to create their own balances, right? Because it is still
39:05
Unidentified speaker (speaker_1): true. I always tell people it's
39:05
Unidentified speaker (UNKNOWN): like it's like
39:06
Unidentified speaker (speaker_1): before you come at me, just go through the thousand steps of actually
39:09
Unidentified speaker (UNKNOWN): building a balance and let me know
39:11
Unidentified speaker (speaker_1): which steps you disagree on. Because I we could
39:13
Unidentified speaker (UNKNOWN): be
39:13
Unidentified speaker (speaker_1): aligned on nine hundred 998 and
39:15
Unidentified speaker (speaker_1): of them.
39:16
Unidentified speaker (speaker_3): So the so so the the the marketplace for ideas does not have enough ideas still, right? And and that's because it's difficult and sometimes expensive to do so, right? So
39:30
Unidentified speaker (speaker_3): I think with
39:32
Unidentified speaker (speaker_3): this is a benefit, this is one thing that AR will help with, we will have more people doing more of that work because like what Keshav Keish mentioned. have the the 80, the 80% that is actually legitimately undifferentiated can can collapse and and people can focus on the 20%, let's say.
39:52
Unidentified speaker (speaker_3): This is in the in the you know the 20% that really is important. And then they can come up with their, you know, express their own opinions, come up with their own views, and and then with more ideas, you're you I think you'll have more. competition and ultimately you end up with better ideas, right? So so that's why you need small vendors. That's why you need large vendors. That's why you need people doing it themselves.
40:12
Unidentified speaker (speaker_3): You need you need everything. So I think that's one that's one kind
40:16
Unidentified speaker (UNKNOWN): of
40:16
Unidentified speaker (speaker_3): So
40:17
Unidentified speaker (speaker_1): let me
40:17
Unidentified speaker (UNKNOWN): let me drill into that
40:17
Unidentified speaker (speaker_1): for one second. Before we go, because I'm curious to hear uh Kishov's views
40:21
Unidentified speaker (UNKNOWN): as
40:21
Unidentified speaker (speaker_1): well on this. But I'm just curious on when you talk about like more and more competition, the belief there is, like I think like the mental model is that
40:28
Unidentified speaker (UNKNOWN): we kind
40:29
Unidentified speaker (speaker_1): of whittle our way towards some kind of consensus truth in the middle. That the process ends when like complete truth is known. Which I think is true if you view truth as being the thing we're optimizing towards.
40:43
Unidentified speaker (speaker_3): Yeah.
40:45
Unidentified speaker (speaker_1): Versus, I think in many cases, as we've seen, you know, in many cases, if if humans remain the customers, truth is not always the thing we're optimizing towards.
40:55
Unidentified speaker (UNKNOWN): You know, I think like
40:57
Unidentified speaker (UNKNOWN): you know,
40:57
Unidentified speaker (speaker_1): you know,
40:58
Unidentified speaker (UNKNOWN): like
40:58
Unidentified speaker (speaker_1): s like let's say someone got on a podcast and said oil prices go to $200. Not that I have any experience with this personally.
41:05
Unidentified speaker (UNKNOWN): Yeah.
41:05
Unidentified speaker (speaker_1): Um, but let's say it happened. Like that gets people's attention in
41:09
Unidentified speaker (UNKNOWN): a
41:09
Unidentified speaker (speaker_1): way, and I think it, you know, in these attentional markets,
41:13
Unidentified speaker (UNKNOWN): I
41:13
Unidentified speaker (speaker_1): mean, we know many people. Like, like you're always these questions like, are you a really bad analyst or are you just a really good grifter?
41:19
Unidentified speaker (UNKNOWN): Um and
41:20
Unidentified speaker (speaker_1): I do wonder, like, does that process, is that process supercharged here? Or do you think because again, like
41:27
Unidentified speaker (UNKNOWN): again, I'm I'm I'm
41:28
Unidentified speaker (speaker_1): I'm going off here, but
41:29
Unidentified speaker (UNKNOWN): I think
41:29
Unidentified speaker (speaker_1): that like one of these ideas, like one of the things even before AI, and we had proliferation of information, we had proliferation of social media. Um, one of the things we talked about was that we we increasingly lost the single point of truth. Like there was no longer the Walter Kronkheid. There was
41:44
Unidentified speaker (UNKNOWN): no
41:44
Unidentified speaker (speaker_1): longer this kind of consensus view of reality that we were kind of going into these different camps of reality. The same thing I think has happened in large part in oil analytics and market analytics generally, whether or not that's currency rates or, you know, you know, heck, uh fixed income. Oh dear lord. Um but like I think all of this stuff and oil, I think, is a great example.
42:03
Unidentified speaker (speaker_1): In some ways, the part the camps have actually never been further apart than they are today. Which again almost runs counter to what we would think the gravitational pull of these of these systems would be. Uh Kishav, it looks like
42:17
Unidentified speaker (UNKNOWN): you have You're
42:18
Unidentified speaker (speaker_1): itching, I love
42:19
Unidentified speaker (UNKNOWN): it. No, no,
42:19
Unidentified speaker (speaker_2): I'm itching. No, no, also like, you know, when we started the industry to now, the the composition composition of market participants have changed dramatically. So you've got the high frequency traders, and you know, I mean, there's no way to like
42:35
Unidentified speaker (speaker_2): document this, but the percentage of players who are physically involved in the oil market, my guess is as a percentage, that's just been a declining slope in the oil markets. And the percentage of high frequency traders
42:48
Unidentified speaker (UNKNOWN): like
42:48
Unidentified speaker (speaker_1): the virtual barrel, like Ilya uh Bushuev. who we've had in the show, um, was it like a thousand a hundred to one, a thousand to one? Like these numbers are increasingly bonkers different.
42:57
Unidentified speaker (speaker_2): Yeah. And and it's also reducing the importance of balances because what is the point of doing balances if no one cares about them? You know?
43:06
Unidentified speaker (speaker_2): To put it very bluntly, you know? So my point is like, what make as to question what makes the market? You know, it's it's the market participants, and you can even see an open interest. You can see the open interest in futures going down.
43:16
Unidentified speaker (speaker_1): Yeah.
43:16
Unidentified speaker (speaker_2): Um, and the open interest in options is dramatically going up. Because with higher volatility and it's a self-fulfilling loop, if you got higher volatility, you don't want to be touching futures. You want to be going on options. Um,
43:30
Unidentified speaker (speaker_2): firstly, like where we if you think like where AI is heading, you
43:34
Unidentified speaker (UNKNOWN): know,
43:34
Unidentified speaker (speaker_2): or or if this keeps turbocharged, it's very hard to predict where AI will be in six months' time. versus what three years because even this time last year, what we have done as a company, I thought was unthinkable, like having information on demand.
43:48
Unidentified speaker (speaker_2): And like
43:50
Unidentified speaker (speaker_2): like we're never gonna get perfect information because the perfect cat and mouse game. The market's just gonna get more and more complex. And we haven't even like talked about how oil is linked with macro and metals. And it's just it's an impossible game. to get perfect information. And you know, Neil and I are the most AI pilled people, like you
44:08
Unidentified speaker (UNKNOWN): know,
44:08
Unidentified speaker (speaker_2): we are very AI positive, but we gotta talk about the limitations of AI. So AI or in in any market is a giant consensus machine.
44:17
Unidentified speaker (speaker_1): Yeah.
44:17
Unidentified speaker (speaker_2): So AI will give you consensus in any given market.
44:22
Unidentified speaker (UNKNOWN): It's it's the
44:22
Unidentified speaker (speaker_1): best midwit, right? I think is the kind of
44:25
Unidentified speaker (speaker_2): and and money, money. is always made when you deviate away from the consensus. So I don't like, you know, and Neil and I have differ different opinions on how AI can create alpha. You know, there's a lot of work to be done where how AI can help you. Uh right now we're at the stage where AI can optimize a lot of your life, but when it comes to
44:45
Unidentified speaker (speaker_2): generating alpha or pure returns, you know, you still need a lot of human judgment. And even in a perfect market, or perfect information, that will still continue and evolve. It's a long-winded answer, but you know, if you have
44:58
Unidentified speaker (UNKNOWN): a lot of people.
45:06
Unidentified speaker (speaker_1): That I always think that like if someone gets lucky and uh if a person gets lucky a bunch of times, it's it's read as luck. Whereas I think we're much more capable. of like saying like the machine was 55% correct uh or whatever. Like I think we're more comfortable ascribing probabilities to machines than we are to people. Uh because again, I
45:26
Unidentified speaker (speaker_1): think people at the end of the day are actually not we like to we like to we like to talk about probabilities. Like people are generally very bad at thinking natively probabilistically. We're like we're far more causal. Um
45:37
Unidentified speaker (UNKNOWN): and
45:37
Unidentified speaker (speaker_1): again, even then, I
45:38
Unidentified speaker (UNKNOWN): think
45:38
Unidentified speaker (speaker_1): like like the balances, like I agree. Like I I completely agree with what you're saying. But I will also say that P the P people will always say like balances don't matter.
45:47
Unidentified speaker (UNKNOWN): And
45:47
Unidentified speaker (speaker_1): then the next day they'll be out talking about how bullish or bearish a particular balance is that confirms their bias. Like it's it's like the perfect, like it's the perfect quote. It's like uh, you know, we're all it's it was the the the the you know, the Keynes quote, like we're all like uh, you know, uh subject to some. kind of defunct barrel counter at the end of the day. Like like at the end of the day,
46:07
Unidentified speaker (speaker_1): people will still use this stuff,
46:09
Unidentified speaker (speaker_0): but
46:09
Unidentified speaker (speaker_1): only if it agrees with whatever they wanted it to.
46:12
Unidentified speaker (speaker_0): Yeah.
46:12
Unidentified speaker (UNKNOWN): And
46:13
Unidentified speaker (speaker_0): all all
46:13
Unidentified speaker (UNKNOWN): models are
46:13
Unidentified speaker (speaker_0): wrong, but some
46:14
Unidentified speaker (speaker_3): are useful.
46:15
Unidentified speaker (UNKNOWN): Yes, is a good quote.
46:16
Unidentified speaker (speaker_2): And and and human bias is so strong. And this is what something we we tell people like, you know, you can use AI to counter your human bias. So you know, if you're reading a PDF, you know, um, you could you could assess what the the biaship is. And even if with AI, the technology is already there, but where the
46:36
Unidentified speaker (speaker_2): big challenges, how do you overcome human biases and also general like your existing habits? So that's once people are more aware of it, then you can use that to neutralize your existing biases or whatever.
46:52
Unidentified speaker (speaker_2): Not bad habits, but you know, just just the way you
46:56
Unidentified speaker (speaker_2): come around. A lot of people think that the the way they think is they have a few ideas and a few data sets and a few reference books. And that's how they come with the view.
47:06
Unidentified speaker (UNKNOWN): I
47:06
Unidentified speaker (speaker_1): want to I so I wanted to ask,
47:08
Unidentified speaker (UNKNOWN): and
47:08
Unidentified speaker (speaker_1): this might be and and uh let me know if this is if this is even cool. But uh I want to try an experiment. Um I'm curious if if you guys have like the like the platform open and can test a query. Because one of the things I'm curious about is this idea of like where where the water's edge of intelligence is, like where and we
47:28
Unidentified speaker (speaker_1): were talking about this a little bit in the but before we turned on, before we started recording, but this idea of like in this current moment, the question is in the market, like all questions lead back to like. what will happen with the Iran war through the balance of this year and into 2027. And very frankly, I'm pretty
47:48
Unidentified speaker (speaker_1): confident that Donald Trump doesn't know what he's going to do with the war. So I don't know how enhanced, I don't know where the value of intel uh how far enhanced or even perfect intelligence can get us into answering. that question when the principal himself doesn't really kind of know. So, like I
48:08
Unidentified speaker (speaker_1): I posted a um
48:11
Unidentified speaker (speaker_1): a poll on on Twitter uh earlier this morning, basically like, does the war, does the war end with a US deal? Does it, you know, do we get material re-escalation? Is there some kind of what I call the unilateral US taco? Or do we just get like a muddled through messy status quo over the next year? I'd be curious: like, how would your system
48:31
Unidentified speaker (speaker_1): answer that question?
48:34
Unidentified speaker (UNKNOWN): And
48:35
Unidentified speaker (speaker_1): like, how would it even talk about it? Because I think again, this is the type of things where it's like,
48:39
Unidentified speaker (UNKNOWN): I
48:40
Unidentified speaker (speaker_1): don't know the answer. I think I I'm actually very confident I do not know the answer. Um, I'm very confident that no one knows the answer. But in these moments where it becomes this kind of parameter. parameterization, like how how would these systems answer these impossible questions?
48:56
Unidentified speaker (speaker_3): Yeah, I mean, I I I guess I guess I would start with saying I agree agree with you. I think that I don't think that there is anybody who knows the answer, right? Um I mean Kishan, if you want to run the the run it and see, I'd I'd be curious.
49:08
Unidentified speaker (UNKNOWN): I'm curious,
49:08
Unidentified speaker (speaker_1): I'm curious.
49:09
Unidentified speaker (UNKNOWN): I think The system it
49:10
Unidentified speaker (speaker_3): is is guided to not like
49:14
Unidentified speaker (speaker_3): emptily pontificate.
49:20
Unidentified speaker (speaker_3): Yeah. So, you know, this case just running, it it's going to try to use a synthesis of available news as well as available data. And and so one thing that it leans on is like polymarket, for example. Um so yeah.
49:36
Unidentified speaker (UNKNOWN): Okay, interesting. So so it would it
49:38
Unidentified speaker (speaker_1): would kind of interpret. polymarket probabilities as some almost like a
49:41
Unidentified speaker (UNKNOWN): polling.
49:42
Unidentified speaker (speaker_3): It's it's gonna prefer like I mean if you push it it it will it will
49:47
Unidentified speaker (UNKNOWN): you
49:47
Unidentified speaker (speaker_3): know espouse an opinion, right? But like I think I think it'll it'll try to lean on lean on like something more more concrete. Yeah.
49:56
Unidentified speaker (speaker_2): And and and really like with with these AI stuff, the first answer is not is never the most useful
50:02
Unidentified speaker (speaker_1): output.
50:03
Unidentified speaker (UNKNOWN): It's when
50:03
Unidentified speaker (speaker_2): you go
50:04
Unidentified speaker (UNKNOWN): down the rabbit hole.
50:05
Unidentified speaker (speaker_3): It's the iteration. I think I think I think it's the the brainstorming aspect. Yeah. It's the the mental gym gym, really. And and and you know, like one thing that I get a little bit sad about is like
50:18
Unidentified speaker (UNKNOWN): it
50:18
Unidentified speaker (speaker_3): seems like the main use case for AI systems right now is to generate large volumes of shallow work, right? And and I I
50:27
Unidentified speaker (speaker_1): hadn't noticed like
50:28
Unidentified speaker (UNKNOWN): LinkedIn posts.
50:29
Unidentified speaker (speaker_3): Right. And
50:29
Unidentified speaker (UNKNOWN): I and I guess a little more like
50:31
Unidentified speaker (speaker_3): ro romantically as someone who likes technology thinks there thinks that there's a tremendous opportunity for actually for for for depth in in these systems, right? And again, like I mentioned, a little bit of a mirror, right? If you have a hundred ideas, you can get interesting feedback on them versus, you know, I'm just gonna spam
50:51
Unidentified speaker (UNKNOWN): a
50:51
Unidentified speaker (speaker_3): thousand uh uh LinkedIn comments or whatever the case might be.
50:55
Unidentified speaker (speaker_2): Yeah.
50:56
Unidentified speaker (speaker_3): So
50:56
Unidentified speaker (speaker_2): what
50:57
Unidentified speaker (speaker_3): do we got?
50:58
Unidentified speaker (UNKNOWN): So
50:59
Unidentified speaker (speaker_2): okay. So I asked, I didn't know it was the best starting question, but I said give me scenarios on how the US-Iran conflict can end. Okay.
51:06
Unidentified speaker (UNKNOWN): All right,
51:07
Unidentified speaker (speaker_1): we've got we've got Iran's terms, US terms, US policies, and president. Well I think that's that's a it's a it's a good framework to start.
51:14
Unidentified speaker (speaker_2): Okay. And where the markets are currently. So it's gone for for pre-war, uh peak, uh MOU, uh sell-off, and then October average and where we are now. And it's gone for flat price and structure.
51:29
Unidentified speaker (UNKNOWN): Because
51:30
Unidentified speaker (speaker_2): you know, in our market, people care more about structure than flat price. Um and then yeah, polymarket, it's given liquidity, and then different scenarios. It
51:38
Unidentified speaker (UNKNOWN): can
51:38
Unidentified speaker (speaker_1): gives a it g it polymarkets giving a forty-five percent probability that we have a ceasefire within the next six days. That's shocking to me. Some free
51:47
Unidentified speaker (speaker_2): money to be made. That
51:48
Unidentified speaker (speaker_1): seems shocking to me, but okay, I guess.
51:51
Unidentified speaker (speaker_1): Not naturally. And I will
51:53
Unidentified speaker (UNKNOWN): eat my words.
51:54
Unidentified speaker (speaker_1): By the time this is published, we will actually be in a ceasefire and I
51:57
Unidentified speaker (UNKNOWN): will
51:57
Unidentified speaker (speaker_1): add a I
51:57
Unidentified speaker (UNKNOWN): will
51:57
Unidentified speaker (speaker_1): automate again. I don't know what's gonna happen. Exactly. Okay, so this is this is good. I think this is management. So basically, a deal. Um, so basically it's giving, let's say, a deal and a muddling through, kind of a 70% odds. And then escalate. Interesting. Okay.
52:15
Unidentified speaker (UNKNOWN): So
52:15
Unidentified speaker (speaker_1): what's funny? So this it's it actually come came to, I think, the the same rough uh kind of scenarios without even guidance. I think the same rough scenarios that I had kind of um kind of enumerated.
52:27
Unidentified speaker (speaker_1): Cool. Um and it's giving,
52:29
Unidentified speaker (UNKNOWN): I think,
52:30
Unidentified speaker (speaker_1): reasonable probabilities, actually. Um I probably would add more to the muddle through than the than the deal personally. Um and I definitely think the market right now has given a higher, or at least the market was giving a notably higher probability to escalation than 15% here. But I think that this is this is a you is a useful framework. So I think this is actually
52:50
Unidentified speaker (speaker_1): like I would say this is. a reasonably successful uh kind of demonstration, honestly. I would need to dig through everything and I would obviously have problems with everything, as I do with everything.
52:58
Unidentified speaker (UNKNOWN): But that's what
52:59
Unidentified speaker (speaker_3): that's what
52:59
Unidentified speaker (UNKNOWN): I'm saying.
53:00
Unidentified speaker (speaker_1): Yeah. Exactly.
53:02
Unidentified speaker (UNKNOWN): Yeah.
53:02
Unidentified speaker (speaker_2): And and Rory, like, you know, like a a lot of market participants, you know, they think history repeats itself. You know, so there'll be a big chunk of the market who'll think, oh, are we at are we in MOU, MOU 2.0? Where there'll be a lot of optimism, prices sell off. and then we're back to square one. So now what the platform does is that you can see how uh the price reaction was
53:22
Unidentified speaker (speaker_2): during MoU and when MOU fell apart. So if you have that view that this is another uh head fake, you know, then you can put on certain positions and assume that you're gonna get the same price action that you got before MOU and post. So this is where it gets very powerful This if you have a narrative that, oh, this is gonna get dragged out again,
53:42
Unidentified speaker (speaker_2): you have all the recent history for you to come up with a better um answer.
53:48
Unidentified speaker (speaker_1): Okay. So we're getting along in the truth in this interview, but I want to end with like one final question for you both, which is
53:56
Unidentified speaker (UNKNOWN): again,
53:57
Unidentified speaker (speaker_1): going back to say like
53:59
Unidentified speaker (speaker_1): completely acknowledging that it's impossible to know. Forward, you put yourself forward two years. What will be different, like fundamentally almost like unidentifiably different from the job of an oil market, an oil market analyst or oil market observer in two years, and
54:19
Unidentified speaker (speaker_1): say, let's say the beginning of 2029? twenty twenty-nine.
54:24
Unidentified speaker (UNKNOWN): Um, what
54:25
Unidentified speaker (speaker_1): does that look like versus? now? I'll I'll start with Kishov.
54:29
Unidentified speaker (speaker_2): Um
54:30
Unidentified speaker (UNKNOWN): I
54:30
Unidentified speaker (speaker_2): mean like you know with
54:31
Unidentified speaker (speaker_1): predict predicting the future is very tricky but
54:34
Unidentified speaker (UNKNOWN): we'll be
54:34
Unidentified speaker (speaker_1): wrong but usefully so hopefully
54:35
Unidentified speaker (UNKNOWN): no
54:36
Unidentified speaker (speaker_2): I think I I think where you
54:39
Unidentified speaker (UNKNOWN): know
54:39
Unidentified speaker (speaker_2): they talk about the K-shaped economy in in macro, you know, where the people, uh if you're asset owners, you're doing really well. But if you're like on on on a on a wage you're you're not doing as well. Similarly,
54:50
Unidentified speaker (UNKNOWN): in
54:50
Unidentified speaker (speaker_2): in AI we are already seeing people who are using it daily versus people who are like oh ai can't do this because I have 10 years 15 years 20 years of experience or they're not using it daily because for whatever internal limitations that they have yeah but people who are using it daily the compounding that's going on so you're gonna get a two like another K a
55:09
Unidentified speaker (speaker_2): K-shaped uh path where you're gonna get a super analyst or a super trader where
55:17
Unidentified speaker (speaker_2): you could just do everything by by yourself or or or or just the productivity is like a thousand X. Um so we but we will still see a lot of people who the divergence will will grow. Um but in terms of capabilities, um
55:36
Unidentified speaker (speaker_2): you could be a one one-person, person, one one-man band and do everything, and nothing is impossible, you know.
55:45
Unidentified speaker (UNKNOWN): Well,
55:45
Unidentified speaker (speaker_1): you're speaking my dream there, man. I don't want to turn into an HR manager.
55:50
Unidentified speaker (speaker_2): But
55:52
Unidentified speaker (speaker_2): but but but human judgment. So if you have domain expertise, like you know, like you and I, you know, Neil, we we run our own businesses,
56:01
Unidentified speaker (speaker_2): you've got
56:03
Unidentified speaker (UNKNOWN): what
56:03
Unidentified speaker (speaker_2): do you call if you've got domain expertise, you this is super, super powerful. So um but but human judgment is gonna be more and more important because you need to figure out where AI is not valuable and where it's valuable.
56:18
Unidentified speaker (speaker_2): That's that's the yeah.
56:20
Unidentified speaker (UNKNOWN): All
56:20
Unidentified speaker (speaker_1): right, and Neil?
56:21
Unidentified speaker (speaker_3): Yeah, I mean I I think
56:22
Unidentified speaker (UNKNOWN): I
56:22
Unidentified speaker (speaker_3): think if we're if we're gonna take like if you're maybe an analyst, um I think you can look to what's happened in software as a as a roadmap for one, two years. So the transformation over the last two years for like a a software engineer or data scientist is tremendous. You know, we've gone from typing every line of
56:42
Unidentified speaker (speaker_3): code to uh prompting but then reviewing every line of code that the agent gives back to now you know, we're um most many programmers are launching five, six, seven agents to work on things and
57:01
Unidentified speaker (speaker_3): designing uh systems to give those agents feedback on if what they did was correct or wrong versus checking every number. Um so in software, if that's like if this runs, send me a short. video of how it works, right? And and that's a good way to check if it works or not. Maybe for an analyst, there's some
57:21
Unidentified speaker (speaker_3): invariance you know about the outputs, right? That that used to correct to check. So so the, you know, that that that I think is one thing. And then and then like Gish have mentioned, what an individual can do is just massively amplified. So when I started programming, there was this concept of a 10x engineer. There was one controversial
57:41
Unidentified speaker (speaker_3): concept that, like, on every team, there was one person who did time 10 times as much as everybody else. And it feels like that's going to 100x or 1000x, right? Um
57:54
Unidentified speaker (UNKNOWN): the
57:54
Unidentified speaker (speaker_3): other thing I'd say about software is that it's it's an incredibly interesting time because everything feels fresh. And I think that's how it will feel for a lot of analysts. And it's also a little bit of like an uncertain time because
58:11
Unidentified speaker (speaker_3): the thing that you relied on is your kind of core competency is just is is so so so different. Um
58:19
Unidentified speaker (UNKNOWN): Yeah,
58:20
Unidentified speaker (speaker_1): I don't know if it's
58:20
Unidentified speaker (UNKNOWN): I don't know if it's
58:21
Unidentified speaker (speaker_1): I don't know if it's more more apprehensive to be like a middle career analyst or an early career analyst, right? now. I actually legitimately don't know which one is
58:29
Unidentified speaker (UNKNOWN): ours.
58:29
Unidentified speaker (speaker_3): I would try to take
58:31
Unidentified speaker (UNKNOWN): a
58:31
Unidentified speaker (speaker_2): whole new podcast for
58:33
Unidentified speaker (UNKNOWN): it. I would
58:33
Unidentified speaker (speaker_3): take, I would kind of take the optimistic point of view, right? That if if you as an individual are able to produce more, that should mean that you are able to negotiate more for yourself and there should be more demand for your service, right? You know, like the sort of like Jebins paradox kind of. view. Um
58:53
Unidentified speaker (speaker_3): But I I think I think what is happening in sop software is is unique in that the the work is very easy to automate by these systems and every other type of work is harder to automate than software for a variety of reasons. Um, but I think it's a good roadmap for technical skills, you know, uh technical skills that are already entirely or
59:12
Unidentified speaker (speaker_3): mostly on the computer.
59:14
Unidentified speaker (speaker_1): All right. Well, Kishav and Neil, thank you so much for joining us in the All Ground But Podcast. Before I let you go, could you just let our audience know where to find you guys uh and find more of your work?
59:25
Unidentified speaker (speaker_2): Thank you very much, Roy, for having us. But yeah, the best way you can find us is hiloanalytics.com
59:30
Unidentified speaker (UNKNOWN): or
59:31
Unidentified speaker (speaker_2): you can contact me on LinkedIn, you
59:32
Unidentified speaker (speaker_0): know,
59:32
Unidentified speaker (speaker_2): all Oh the usual
59:34
Unidentified speaker (speaker_1): We could can we we we we can bully them into tweeting more.
59:37
Unidentified speaker (speaker_2): Oh
59:37
Unidentified speaker (UNKNOWN): yeah, yeah,
59:38
Unidentified speaker (speaker_2): yeah, yeah.
59:38
Unidentified speaker (UNKNOWN): Please yeah. Yeah. I
59:40
Unidentified speaker (speaker_2): w I wish
59:40
Unidentified speaker (UNKNOWN): you
59:40
Unidentified speaker (speaker_2): know, I wish I was a celebrity like you, Rory, on on on Twitter. I'm I'm I'm very impressed.
59:46
Unidentified speaker (speaker_1): Well, thank you so much for joining us guys and uh and and we'll have to check in a couple years to see if uh see if you've built uh my Lieutenant Commander oil data yet.
59:55
Unidentified speaker (UNKNOWN): Thanks
59:55
Unidentified speaker (speaker_3): so much. Thanks
59:56
Unidentified speaker (UNKNOWN): a lot guys.
59:57
Unidentified speaker (UNKNOWN): Thank you very much,
59:57
Unidentified speaker (speaker_2): Rory.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights