i'm joining the team at @ExaAILabs!
search is one of the most important pieces of the entire AI stack and Exa is building the best search engine in the world.
excited to get to work!
Charles retweeted
Exa now indexes 3,000,000+ podcasts.
Your agents can now search across millions of hours of transcripts, updated daily, and track what the world is talking about.
Charles retweeted
We're excited to announce Exa 3.0!
A huge number of improvements:
- Bigger, better embeddings model driving core search
- Higher quality highlights with further reduced token counts
- Lower latency, cheaper search type "Lite"
- Better reranking, particularly across long documents
- Even better real-time news, sports, weather, and market data
- Laundry list of other stuff
Let us know what you think! 🫡
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 🧵
Charles retweeted
Exa has released Exa 3.0, scoring 80 on the Artificial Analysis Search Index in Dynamic Highlights mode, tying with Perplexity Search for the #1 spot
@ExaAILabs has released a new Dynamic Highlights mode to its Search API. Given a search query and its results, Dynamic Highlights determines how much text to return from each page within a shared budget
We benchmarked two Exa search types with Dynamic Highlights ahead of release: an upgraded auto mode, which balances search quality and speed, and a new lite mode, optimized for lower cost and latency. Both were evaluated using our Stirrup agent harness and GPT-5.6 Luna (medium) as the agent model
Key benchmarking results for Exa Search with Dynamic Highlights:
➤ Ties for #1 on the Artificial Analysis Search Index: Exa Search (auto, Dynamic Highlights) scores 80, level with Perplexity Search (medium) in the #1 spot, while Exa Search (lite, Dynamic Highlights) scores 79. On DeepSearchQA, auto (85.6 F1) and lite (84.9 F1) rank 1st and 2nd respectively, 4 to 5 points ahead of Perplexity Search (medium, 80.8)
➤ Reduces search calls and token usage: Exa Search (auto, Dynamic Highlights) uses 7.6 searches per task, the fewest of any Search API on the leaderboard, down from 9.4 with standard highlights. Tokens per task fall 47%, from 299k to 157k
➤ Reduces cost per task by 21%: Exa Search (auto, Dynamic Highlights) costs $0.066 per task, compared to $0.083 with standard highlights. Lite costs $0.044 per task, with both sitting on the cost vs. capability Pareto frontier. Exa Search is priced at $7 per 1k searches for auto and $4 per 1k searches for lite
Charles retweeted
Exclusive: Instinct is using Exa for web search
We gave Instinct URLs for nonexistent pages on a website we administer. Since the requests were for fake pages and the timing matched our conversation with Instinct, the requests couldn’t plausibly have come from anything else.
In total, it made about 12 requests over an hour and a half. We checked Cloudflare analytics and the server’s raw request logs to get the full picture.
A few things stood out:
- Instinct used Exa to fetch pages. Every direct fetch arrived with the user agent Mozilla/5.0 (compatible; ExaUser/1.0; +crawler.exa.ai/exauser), from a different IP each time.
- Exa signs its requests (Web Bot Auth, Signature-Agent: "crawler.exa.ai"). But nothing in the requests identifies Instinct. They are indistinguishable from those of any other product built on Exa.
- In one test, after the fetch returned a 404, Instinct loaded the page in its browser without being asked.
- Instinct’s browser had no self-identifying features. It looked like standard Chrome on desktop Linux from a residential IP address, with no bot user agent or signature.
We directly asked it to use the browser in two other tests. All three browser visits came from the same residential IP over about 80 minutes. The browser kept its cache between visits, consistent with a persistent browser behind a residential proxy. We can’t say whether that IP is stable over longer periods or shared with other users, so it’s not a reliable way to identify Instinct.
This presents unique challenges for marketers hoping to track Instinct traffic to their site. Unlike ChatGPT, Claude, or Meta, Instinct doesn’t send its own user agent or signature. Its fetches look like generic Exa traffic, and its browser visits look like a normal user’s.
All of this means you can’t track, allow, or block Instinct specifically. The only reason we could link these requests to Instinct was that we controlled the URLs and the timing. For normal traffic, you don’t have those luxuries.
Link to it on sunbeam: sunbeam.ai/thoughts/instinct…
Charles retweeted
10x a year for 3 years? I've seen a lot of growth, and that is truly exceptional. It's not unusual for startups to grow 10x in a year, but to keep it up for 3 years is close to unheard of. That's 1000x.
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 🫡
Charles 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 🫡
Charles retweeted
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
Charles retweeted
If you think LLM benchmarks are a shitshow, you should learn about the web search benchmarks space 😅
ATLAS is a solid step towards advancing it!
We built ATLAS, a web search benchmark for how agents actually search today.
Existing agentic search benchmarks are often outdated, contaminated, or saturated. ATLAS is designed to separate strong search from weak search.
Learn more: exa.ai/blog/atlas-benchmark
Charles retweeted
We built ATLAS, a web search benchmark for how agents actually search today.
Existing agentic search benchmarks are often outdated, contaminated, or saturated. ATLAS is designed to separate strong search from weak search.
Learn more: exa.ai/blog/atlas-benchmark
Charles 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!
Charles retweeted
We just added 211 million restaurants, museums, and parks to Exa's index.
Give your AI agents access to search over all the places in the world with structured data including opening hours, reviews, and phone numbers.
Charles retweeted
after building agentic design systems for multi-million dollar startups, I want to start sharing everything I’ve learned
starting with free skills for your coding agent:
design.how View article
Charles retweeted
Introducing Agent Ultra - a step change in deep research
Agent Ultra orchestrates swarms of agents to perform exhaustive research, build comprehensive lists, and answer questions requiring thousands of sources.
In both evals and vibes, it's state of the art
Charles retweeted
Jev with Exa is INSANE.
> Jev without websearch confidently gives wrong outputs
> Jev with websearch is literally much more accurate
Try Jev (with Exa websearch) for free 👇
Charles retweeted
We built a time machine for the web.
Introducing Exa Snapshot: an index of 400 billion historical snapshots of webpages that lets you search as if it's the past.
Snapshot is already being used for backtesting prediction models, RL at labs, exploring the pre-AI web, and more.
Charles retweeted
This is the first labor day where you can sense a future where human labor isn't necessary.
But do we even want this future?
I spent much of this weekend thinking about this question. I went through the full cycle of excitement, fear, meaninglessness, back to excitement.
The world could be so amazing? Could we ever control this level of power? But will humans have any value?What do we get if we make it through to the other side?
Things became clearer when I made a pros/cons list. So here's a pros/cons list for the year 2050 where superintelligence occurs.
I assume in 2050 we'll have superintelligence far smarter than all humanity combined. I assume a billion humanoid robots. I assume essentially anything can be built given sufficient compute + energy.
A 2050 future done poorly:
- extreme power concentration in 10 or so tech firms who control compute. Most humans feel powerless, and the same with governments
- Authoritarian government gets AI first, overwrites all digital information, no one knows what's true. This government ensures hegemony for centuries
- major virus with Ebola death rate + as contagious as Covid wipes out much of mankind. Launched by a suicidal terrorist group, it's designed to not be detectable until a year in
- Nuclear war over fear of US establishing permanent superintelligence supremacy. No humans left. Just robots roaming a wasteland earth
- Misaligned, self-replicating AI swarm spreads throughout the universe turning all life into datacenters, in order to answer some impossible question on an eval
- all the dangers we can't imagine: severe digital addiction, dystopian government structures, etc.
A 2050 future done well:
- Pristine health: the cure to all major diseases once and for all. No cancer, Alzheimers, heart disease, etc. Easy to write this, but imagine it actually happening. We've done it before, we can do it for everything.
- Aesthetics: Aging slowed down or reversed. No more acne, hair loss, wrinkles, muscle loss
- Beautiful earth: climate change a thing of the past. Solar + wind + fission + fusion. Super clean air everywhere. Robot ships fully clean the oceans. Forest restoration. All life flourishes.
- Beautiful cities: robots ensure clean streets, very little crime, stunning buildings built in a couple years, way larger houses, super beautiful walking around
- Magical transportation: Jetpacks from house in the woods to the city, hyperloop from nyc to paris in an hour, 5min to anywhere in SF
- Space exploration: space elevators for quick trips to moon, fusion rockets and cryosleep for first trips to Alpha Centauri
- Discovery: Math and physics discovery accelerates exponentially, dark matter turns out to have wild explanation, aliens detected in patterns of galaxies, consciousness finally understood.
- Mental state: drugs without negative side effects help transform the brain into whatever state we want, whether smarter or happier, without any lows or de-motivation. Nothing in physics prevents this.
- ^All these good things but for everyone. In an age of abundance, there's no reason anyone to be left behind. AI helps us figure out government and economic structures that are stable for centuries while enabling individual freedom. New information tools keep society fully informed so we can finally run politics as adults.
- all the good things we can't imagine: communication with animals, new senses, mind uploading into spacecraft, etc.
Don't believe things could get that good? Try explaining our world to a hunter-gatherer from 50,000 BC. Don't believe things could get that bad? Try watching a documentary about 1940 europe.
In my current opinion, the pros of superintelligence outweigh the cons, but only if we tread extremely carefully.
Whether you agree or not, this transition does seem like it's happening. You may choose to slow it down, prevent it, or accelerate it, but we are all now participants in it.
And that brings us back to Labor Day. This transition is going to require an immense amount of human work to go well.
AI labs are providing the initial sparks of the transition, but we need all humans to provide the rest of the infrastructure. From AI tooling to thoughtful policy decisions to manufacturing to media to education. Work's going to change for everyone, but there's never before been more work to do for everyone.
And looking forward decades, no matter how autonomous AI gets, if we have the right societal structures, then we'll still need many humans to steer that future. Imagine, for ex, in 15 years we want NYC to have underground hyperloop transportation. Sure superintelligence can coordinate it and robots can do the actual building, but humans will be needed to handle all the policy decisions and debates over which loops go where, who controls them, who maintains them.
Now multiply that by all the projects we'll want to pursue in parallel across the planet, with limited compute/energy/material resources.
There's plenty to be excited about. There's plenty to fear about. But on this labor day, there's no need to feel meaningless.
There is and always will be many days of labor ahead of us.
Charles retweeted
We're creating Exastential - a magazine to discuss the most existential period in human history, brought about by AI.
Seeking writers and artists to submit piercing essays, videos, whatever perspective you think needs sharing in whatever format.
There's no shortage of topics that need your help exploring. Whole swaths of humanity are currently being overhauled -- education, relationships, societal structure, the economy, war. We all bear the weight of the future, we all should help shape it.
Take education. I have nieces and nephews, and I genuinely have no idea what schooling means anymore in the age of personal AI tutors. Relationships are changing too -- my nieces no longer ask me for help with math homework.
On that note, math could be automated within a year. Physics falls next. Does the human mind matter anymore? I have perspectives on this, and I'll be submitting something too.
Decades are happening in years. Soon they'll happen in months. Help us all become intellectually and emotionally ready. Submissions are open:




















