Great example of how Exa Snapshot can be used to evaluate models without any temporal leakage!
FutureSim Q3 is here!
Opus 5.5 dethrones GPT, displaying strong priors and incredible test-time adaptation.
Opus 5.5 is a beast, calling 50+ subagents and taking over 400 research notes during the 90 day simulation while also being 50% more token-efficient than 6.1 Sol. (1/5)
8
423
Jan van der Vegt retweeted
Exa now runs in Europe.
Our index and retrieval, rebuilt in Stockholm. Requests are stored and processed in the EU.
Available today for enterprise customers.
Excited to be judging this awesome hackathon!
Meet the jury of the SF AI Research Hackathon!
30+ researchers, engineers & founders from @OpenAI, @GoogleDeepMind, @AnthropicAI, @thinkymachines, @Meta, @nvidia, @ExaAILabs and more.
150 hackers. 2 days. Pushing the frontiers of AI research. 🧠
4
236
I stay at home specifically for a delivery and then the delivery person claims I was not at home, and now I have to do this again tomorrow. This is not acceptable @UPS
2
7
594
Jan van der Vegt retweeted
Introducing Exa Monitors V2
Get alerts across the largest index of web + proprietary data
- 130B+ web pages
- 211M businesses
- 1.3B products
- 1B people
Most monitors just rerun search, missing results. Exa runs entity recognition on the whole web. Now 7x cheaper at $2/1k runs
Jan van der Vegt retweeted
Introducing Exa 3.0: our largest ever improvement to search quality.
Exa is now #1 on every benchmark we tested, from public evals to verticals like coding and GTM intelligence.
It's the result of multiple research breakthroughs 🧵
Come watch my talk about Exa Skeleton, our internal data processing framework that manages 100s of billions of pages
A new ranking idea has to work across an index of hundreds of billions of pages.
Go inside @ExaAILabs Skeleton with Jan van der Vegt @DasDzy at #Reverie2026, from ingestion to massive backfills.
Apply to attend → na2.hubs.ly/H08mQ5T0
Jan van der Vegt retweeted
Exa just passed $100m in annualized revenue.
We've grown almost exactly 10x per year for 3 years. We now handle 2 billion requests a month from 200k+ teams. And it’s still so early:
We predict agents will search 1000x more than humans by 2030.
Agents don't search like humans. They write long, complex queries and want all the world's data, not just the web. We've spent 5 years building a search engine designed for them.
We now serve personal agents, AI labs, coding agents, GTM agents, really any agent. Exa gives these agents access to 100s of billions of web documents and tons of proprietary data.
Our customers are increasingly noticing that information is the bottleneck to agent performance. While the world seems close to perfect intelligence with models like Fable and Astra, we are very far from perfect information.
To our small but legendary team, thank you for getting us here, from engineering to research to GTM to ops.
To our customers, it is an honor to serve you. We have some exciting releases coming soon and can't wait to share them 🫡
I love Exa and I love 3Blue1Brown, very cool to see this conversation about the future of search in an agentic world
We sat down with @3blue1brown to discuss our research vision of "Attention over all the world's information."
Plus why knowledge should not live in model weights, the compression trick behind cheaper retrieval, and why we hire through work trials.
1:20 – What is “attention over everything”?
4:46 – A knowledge graph of the whole web
11:32 – How do you evaluate search for AI agents?
15:13 – Why not put everything in the weights?
22:55 – Why Exa hires through work trials
25:15 – The compression trick behind cheaper retrieval
27:40 – Dealing with misinformation
5
377
Jan van der Vegt retweeted
We just added hundreds of millions of places to our index.
For each entity we add, we extract meaning and structure and add to an index of hundreds of billions.
Let us know if there are more types of data you'd like your agents to have access to!
Exa Snapshot allows you to train, simulate and evaluate without temporal leakage while still supporting search.
"Exa is the only external retrieval tool used in this study. Exa, a web index that offers timestamped snapshots, allows us to restrict retrieved information by date."
Jan van der Vegt retweeted
I used Shortcut to cut EPS forecast errors by 20% vs. Wall Street consensus.
I did this by backtesting Q2 2026 earnings forecasts for 113 small caps covered by 3+ analysts on Shortcut with Sol medium effort. Even after raising consensus by 2.7%, the best simple adjustment I found in hindsight, it only cut median errors by 4%, versus Shortcut’s 20%.
The reason I’m exploring this is because I believe that agents will give investors a starting point for companies they might not have previously had time to research. I used Shortcut’s generated spreadsheet models from the backtest to inspect, beyond just the final EPS estimate, which approaches held up. In 40% of forecasts, agents built up to EPS from revenue, costs, taxes, and share count. For that group, median forecast errors were 32% lower than the Street’s.
I also dug around to see how others are approaching earnings prediction. LSEG’s StarMine improves on consensus by giving more weight to recent estimates and historically accurate analysts. A 2024 paper I found reported that GPT-4 outperformed analysts at predicting whether annual earnings would rise or fall, but that was later withdrawn after a co-author found inconsistencies in the data and analysis during replication.
That made how I test just as important as the forecasts. So I realized I needed to limit the agent to information available before each earnings report and account for what it might already know from training.
I chose Q2 reports to reduce the in training-data risk since Sol was trained on data up to mid February. For web search, I restricted access to information from before each report and used TypeSafe’s Jev to screen pages for leaked results, which I found more reliable than Haiku in my tests. I’m now moving to Exa Snapshot to retrieve pages as Exa stored them at the cutoff.
These controls help, but forecasting before results exist is the real test. October earnings season seems like a great time to expand and publish forecasts.
This is still an early experiment, and I’d love to hear what y’all think I should try next and whether this could be useful :)




























