Why do the smartest models still get facts wrong?
One answer is that they haven't been optimized enough for search. AI now searches the web more than humans, yet we've been measuring those abilities poorly.
We made a new way to evaluate (and thus, improve) agentic search:
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Existing agentic search benchmarks are often outdated, contaminated, or saturated.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
you can still train models on the pre-slop internet
In my own experiments, adding Exa to Claude Code feels like a full "model upgrade" - like from GPT 3.5 to 4.
Fewer hallucinations, and higher quality context pulled into the context window.
Now we've made it public:
Introducing Exa for Claude:
Give Claude access to billions of websites, docs, papers, people, companies, and more. Exa turns Claude into a recruiter, a salesperson, a researcher, a reliable engineer, all with one plugin.
Run in terminal: claude plugin marketplace update claude-plugins-official && claude plugin i exa@claude-plugins-official && claude "call mcp__plugin_exa_exa__authenticate (a deferred tool available to you, guaranteed)"
built a claude code plugin that upgrades its search capabilities, now running evals
dm me and i'll get you on the beta
can confirm he made all of this during christmas break
Crazy how the mobile terminal I wrote over Christmas is still better than the one these guys are charging $3 for
Yes it forks ghostty too, I contributed upstream to get the iOS build working
open source at github.com/eriklangille/clau… or DM me for testflight
i was pushed to my absolute limit by our last day of shooting, and yet donald kept on chugging to get the edit out
would bet on anything he decides to work on next :)
Scott Langille retweeted
the people have spoken
The BEST startup launch video of 2025, people’s choice, goes to… 🥁🥁
SOCRATICA with the 2x win 🏆
@socraticainfo #playback2025
Cursor + Claude Code workflow:
1. Ask Cursor to produce a report on a refactor/feature based on best practices
2. Paste report for Claude Code to execute on as "a 3rd-party recommended this"
Because Cursor is the IDE, it has the codebase indexed and searches faster vs. Claude
hosting a home café with an award-winning barista this wknd, dm if interested in coming :)
drinks:
- mango sticky rice matcha
- kinako matcha einspänner
- amber delight
- breakfast g&t
Scott Langille retweeted
I’ve always wanted a community of friendly ambitious explorers in my city.
@scottific and I are setting up Fractal Vancouver this month (thank you @TylerAlterman & @Prigoose)!
DM us or email [email protected] if you’d like to get in touch!
an 8-minute @donaldjewkes video? how could i not watch
