
Eating, gymming, peptides @mixedbreadai
San Francisco, CA
Joined March 2024
- Tweets294
- Following94
- Followers185
- Likes652
agents can do porting pretty well. get inputs, run a source function, read outputs, match byte-for-byte.
we've built and iterated a bunch of internal toolings and libs in this way that could run across diff envs with better performance/quality over the time.
we do have some teammates who just love doing this to see the limit of agents, like randomly ask them to rewrite some libs in rust.
How Phil built a collaborative Word editor with no human engineers.
Inspired by @russelljkaplan's story of how @cognition builds their Android app with just a swarm of Devins.
9
318
Rui retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Rui retweeted
Introducing Expand Search 🔍 Pareto frontier search for your agent.
Two weeks ago we launched Expand Fetch. But which pages should your agent fetch?
We need search. And existing providers are either too expensive or low quality.
Expand Search costs $1.50 per 1,000 searches, with the highest average answer accuracy across the five public benchmarks in our evaluation.
Sign up in the next 24 hours and get 50,000 free credits.
expand the web with lossless scraping
Introducing Expand Fetch, a new way to scrape and crawl the web.
Your agent has never seen the web. No wonder it gets basic questions wrong.
Traditional Markdown APIs lose critical details. So we trained two System one models to change that:
Iris, a vision model that sees pages like you do, preserving tables, icons, and layouts.
X-Ray, a JSON model that uncovers the hidden data behind every page.
Every fetch also comes with a Wayback Machine. Every line links back to its source, so you can verify exactly what your agent saw.
Three products. One API.
On our benchmark of 1,094 questions across 326 challenging web pages, agents using Expand got 74% right. The next best API got 55%.
Available today at $1 per 1,000 fetches.
Try it at expand.ai
1
5
1,736
all you guys gotta pick up some mixedbread merch tomorrow @aiDotEngineer
1
5
157
i think what people still do not understand about code, it is very special kind of workload. almost perfect for ai.
code has an inherent index by being a graph, you can easily traverse everything using the ast etc. unlike most other data you can guess a lot of functions variables etc. by looking up few lines of code.
code is mostly conflict free and the git history and comments give a clear motivation why things happened the way they should have happened.
the models got more persistent, thus they will just spend more tokens to get the data, which is pretty easy for code, for the reasons above. one thing you should not forget as well, cursor did not had the elon gpus, so every token they could saved where good for them and their users.
but most work is unlike code. real world knowledge is massive, full of contradicting, distributed information. you can spend tokens at it, but this is very expensive, doesn't scale and well bottlenecks the models.
as always you want to provide the best possible tools for right job. for coding its grep, for the rest it is not.
Rui retweeted
great opening move makes things way more interesting
1/ We’re excited to share Question’s Gambit module for search agents. ♛
🥇Question Gambit got first on the BrowseComp-Plus leaderboard for recall with BM25 as our only retriever!
📈 It focuses on the first retrieval move in search agents and got 96.6% recall with about 6 times fewer agent search calls.
Same BM25 index. Same agent loop. Just a better opening move.
📄 Paper:
arxiv.org/abs/2609.14412
getting the same level of satisfaction of a good workout from @mixedbreadai search playground
6
211






















