Michael Fine

194 posts

Michael Fine

Michael Fine

@fine_whines

ML @ExaAILabs

Katılım Haziran 2016
2.5K Takip Edilen518 Takipçiler
Michael Fine retweetledi
Exa
Exa@ExaAILabs·
We raised $250M in Series C funding at a $2.2B valuation, led by a16z. Exa is a search lab organizing the web's data for agents.
English
157
169
2K
1.3M
Michael Fine
Michael Fine@fine_whines·
@jld_adriano Better search can naturally lead to better learning in RLVR setups because fewer rollouts totally fail and have zero advantages. But in particular, my hunch is that Exa search degrades more gracefully and has less variance for slightly different query formulations
English
1
0
1
53
Michael Fine
Michael Fine@fine_whines·
Fun nugget from our research here -- using a better search backend during training leads to better performance at inference time not just on the one it was trained on, but also when you swap the backend for each agent.
Michael Fine tweet media
Exa@ExaAILabs

How does Exa compare to Google for training LLMs to search? In this blog post, we find that LLMs using Exa during reinforcement learning reach higher performance with 70% less training compute. exa.ai/blog/rl-search…

English
1
3
13
878
Michael Fine retweetledi
Exa
Exa@ExaAILabs·
Exa powers most of the popular coding agents. We wrote about how we build and evaluate coding-related web search. Blog and open evals: exa.ai/blog/webcode
Exa tweet media
English
12
33
156
75.3K
Michael Fine retweetledi
Exa
Exa@ExaAILabs·
Introducing Exa Deep: putting an agent inside every search For each query, an agent runs in a loop until it gathers all information, then returns structured output. Evals show Deep is Pareto optimal at 4-60s latency, ideal for quick, cost-efficient research!
English
26
54
512
89K
Michael Fine
Michael Fine@fine_whines·
@imbue_ai How does it compare to eg GEPA on prompt optimization?
English
1
0
5
550
Imbue
Imbue@imbue_ai·
Today we’re open sourcing Evolver, a near-universal optimizer for code and text. While benchmarking we achieved SOTA (95%) on ARC-AGI-2 (last week that is 😆) and 3x’d performance of the best open model, reaching GPT-5.2-level performance.
Imbue tweet media
English
41
103
929
128.2K
Michael Fine retweetledi
Exa
Exa@ExaAILabs·
Introducing state-of-the-art People Search: You can now semantically search over 1 billion people using a hybrid retrieval system backed by finetuned Exa embeddings. Try it: exa.ai We also created an eval: exa.ai/blog/people-se…
Exa tweet media
English
130
141
3.1K
3.6M
Michael Fine retweetledi
Exa
Exa@ExaAILabs·
Smarter than your default search engine, faster than your default chat app Try the new exa.ai
English
24
28
336
178.2K
sambhav dixit
sambhav dixit@sambhavdixitpro·
@jeffzwang @fine_whines @feel_ix_ @Will_Bryk @TheIshanGoswami — curious for feedback on v4.3: MS MARCO gains came from [QE + evidence-aware listwise reranker]. Would be great to discuss reproducibility as well as technical details and possible improvements in the future . Really loved how you guys built your entire search architecture , would've loved to do the same if i was in your shoes .
English
1
0
1
169
sambhav dixit
sambhav dixit@sambhavdixitpro·
Not all good ideas are born out of sensible endeavors — some spark from playful curiosity. This weekend i read an article from @ExaAILabs and took up the challenge and so, Project Athena was developed : A novel search stack helping make search simpler , and more effective .
sambhav dixit tweet media
English
1
0
7
331
kalomaze
kalomaze@kalomaze·
this is how the RL reward distribution evolves over time
English
8
9
294
27.8K
Michael Fine
Michael Fine@fine_whines·
We built exa code because we were tired of coding agents burning dozens of tool calls failing to use libraries — give it a try or your codex will end up looking like this:
Michael Fine tweet media
Exa@ExaAILabs

Introducing exa-code, a big step towards eliminating LLM code hallucination. We indexed 1B+ docs pages, Github repos, StackOverflow posts, and more. Given a query, exa-code initiates a hybrid search over this data, chunks it, and returns a concatenated, token-efficient string. In our code hallucination eval, exa-code outperforms every popular web search tool (including Exa!). Available for free: github.com/exa-labs/exa-m… How we built it: exa.ai/blog/exa-code

English
2
1
21
4.9K
Michael Fine retweetledi
Exa
Exa@ExaAILabs·
We raised $85M in Series B funding at a $700M valuation, led by Benchmark. Exa is a research lab building the search engine for AI.
English
197
144
2.2K
1.8M
Vish ✨
Vish ✨@_vish_k·
conic sections
English
1
0
2
148
Michael Fine
Michael Fine@fine_whines·
@_vish_k yesss Vishal finally some coherent and good content
English
1
0
1
47
Vish ✨
Vish ✨@_vish_k·
grice's maxims go hard
English
1
0
1
109
Vish ✨
Vish ✨@_vish_k·
The point of moving to USA is to see the Mac wallpapers irl
English
1
0
2
186
Michael Fine
Michael Fine@fine_whines·
took a lot of eng and ml work getting here! Join us so we can make this ~2 OOM faster and let LLMs search in every forward pass
Will Bryk@WilliamBryk

Exa is now the fastest search API in the world. We built this bc we're a search engine for AIs, and AIs need faster search than humans do -- chatbots, voice agents, etc. This required a crazy amount of optimization across every part of the stack, from training specialized embedding models to designing SIMD optimized data structures in our vector db. It's only possible because we built our own search engine from scratch and so can control every part of the stack. Building it was an insanely hard multi-year effort, but now we're reaping the benefits. Search can be way better than people realize, including speed -- imagine 100ms search 🤯. More to come! Blog post below for more info 👇

English
0
1
7
743
Vish ✨
Vish ✨@_vish_k·
accidentally yelled at a guy on the phone
English
3
0
3
261
will brown
will brown@willccbb·
@lateinteraction this is specifically not what i want to do in this case sorry haha i want an MCP wiki search server (yes, i actually do care that it is a literal MCP server) and i want to hit it millions of times and i want it to cost dozens of dollars, not thousands
English
4
0
6
482
Omar Khattab
Omar Khattab@lateinteraction·
Deep research systems need to discover PageRank 😅 Stop relying on entirely random slop and ancient pages, when there's plenty of official and recent resources....
English
17
16
244
16.2K

Keşfet

@jld_adriano @imbue_ai @sambhavdixitpro @jeffzwang @Will_Bryk @TheIshanGoswami @ExaAILabs @willccbb