agi is here but it got an email job
Joined August 2025
- Tweets4.2K
- Following925
- Followers351
- Likes16.4K
Pinned Tweet
Personalized Arxiv feed
I built a system that allows you to create a personalized feed or search over recent listings. I use a two tower architecture with a preference model and a re-ranker.
uror.io
Fun read exploring credit assignment with agentic judges.
My takeaway: it seems to constrain the reasoning towards what is necessary. Like producing shorter traces with stronger commitment.
Matt retweeted
Very exciting to finally show this off: lossless compression of BF16 models to 12bits (based on NyoomFloat12 from @togethercompute). A lot of work went into this to make it as efficient to run as possible, since no hardware has native support for this dtype.
github.com/dphnAI/nf12
thinking about big model smell again because often I find increased thinking on sol (or small model smell alternatives) increase the surface area of slop.
Instead of evaluating each decision deeper, it will make more "inconsequential" decisions as it relates to the prompt but which often are consequential to the work. It feels like it manifests as a hack to consume thinking while preventing execution drift.
Execution drift in terms of "what it thinks the prompt/prior training is requiring." It'll fixate on what it believes to be the intended goal and does not want to drift from that goal. To do so would increase breadth of search which I bet is a waste of resources in RL.
This is small model smell because often small models need to have greater search space collapse per turn. Otherwise, they don't have the capacity to evaluate the breadth of options.
2
93
TL;DR - Modern LLMs are good judges (comparable to people). Simple prompts of what you want and large models are the ideal judge.
New post in my series "Writing Well": uror.io/posts/writing-well-2…
It's dissapointing but most writing judge studies are bad. I analyze one which is pretty standard in setup and it's claims are not supported.
The vision study I evaluate was better. My reconstruction supports and extends it.
1
6
575
As many neolabs have transitioned to compute providers, prime has gotten worse at compute providing. makes you think
1
176
Matt retweeted
i’ll sweeten the pot.
1k USD prize if you can meet this challenge on a commercial class task, above 0.5B params, in one week.
github.com/DeusExperiments/t…
This always seemed kind of obvious? The issue is you need a converter for every model to every model (which now it's probably feasible due to open router, routing, etc)
MITの研究者らが、AI同士が言葉を使わず頭の中の状態を直接渡して通信する仕組みを作り「テレパシー」と名付けました。
文章でやり取りするより平均40%正確で4倍速かったといいます。
また「画像を見られるAI」が写真の内容を「テキストしか読めないAI」にテレパシーできるそうです。
写真の内容をテレパシーを受信した「テキストしか読めないAI」は、テキストで説明を渡されるよりも写真についての質問に正しく答えられたとのこと。
なお、別の会社のAI同士でもテレパシーできるようにしたのが今回の研究の一番の新規性のようです。
例：
Meta（Llama 3）→ Alibaba（Qwen 3）
Alibaba（Qwen 3）→ Google（Gemma 2）
Google（Gemma 3）→ IBM（Granite 3）
Google（Gemma 3）→ Meta（Llama 3）
Mistral AI（Ministral 3）→ Meta（Llama 3.2）
Google（Gemma 2）→ Mistral AI（Ministral 3）
IBM（Granite 3）→ Hugging Face（SmolLM3）
1
9
489
It's funny that eating bugs was implied to be like physically eating bugs and not the ever growing hunt for bugs in code using human brains to power ai. You will live in the pods (matrix) and you will eat the bugs (phpshell) and you will be happy
1
2
135
Sol 6.1 is the type of guy that's incredibly capable but doesn't listen so the work it does is useless because of stupid decisions it made from not listening
2
86
The goal of @typesafeai Jev model is to unlock Jevon's paradox. I wanted to play around with decision models so I trained TERN-35B-A3B. The goal is price frontier.
On an H100 PCIE I got it to $0.0116/Mtok. Modifications got me to cost-score pareto frontier at $0.00773/Mtok
The model could be a lot better. It was undertrained and I have a lot more data to feed it. The next steps are to improve the data and play with some Deepseek inspired architectural changes. If you want to play around with it in the meantime:
HF - huggingface.co/matthewagi/Te…
Github - github.com/Matthew-agi/Tern
1
35
One final note: Speeds used optimized batch sizes. That likely doesn't reflect the reality of serving but does reflect the reality of serving for RL as a judge and some other tasks.
H100 assumed a price of $2.35 which is what I was renting off @PrimeIntellect
1
18