Matt (@Matthewagi) profile picture

agi is here but it got an email job

Joined August 2025
Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)Photo by Matt (@Matthewagi)
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
Got a nice kick in the butt yesterday to do something I've been putting off. Re-ranker for Arxiv articles is trained and ~industry standard
3
1
1
22
4,271
As was shown with supply chains during covid, there is a lot of modern infra which is incredibly interconnected. Later in this thread everything appears to be traced back to Chase
There is a developing situation with some U.S. banks. This is not yet confirmed as a cyberattack. But it does seem to be spreading. It started with Chase and Bank of America and now Wells Fargo, Zelle, Capital One, and U.S. Bank are also reporting issues. Logins are down on both web and app, and even some ATMs have stopped working. Transfers are down, and some accounts are showing zero balances. There is no official explanation or statement from any of the affected banks yet. There is nothing confirmed that says this is serious yet or has anything to do with AI or hacking. If it develops beyond this point, I will make a thread.
1
89
Just open sourced a dataset of AI generated text by filing within the SEC Text Corpus (350bn tokens).
2
1
1
9
1,156
Every single ML arch argument comes down to quanta of compute rnn vs transformer Chungus vs TTC Looped vs deep
62
Antipasto but it's an old timey gang that's racist against italians
64
Vibe analysis of Part-of-speech break down of comedians vs LLMs (anecdotally this is similar to other writing). The words are more diverse but the grammar is less diverse
@max_spero_ i wonder if yall have ever studied this. like at the aggregate level i would imagine humans disobey formal linguistic rules far more than LLMs, in fact that might be why LLMs have such a distinct smell to them
1
1
9
357
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.
New blog! This one covers my experiments on fine-grained credit assignment with an agentic judge, and how it affects the models' performance and reasoning quality and some thoughts on RL science. Link below, please do check it out, and RT/QT if you like it:)
3
1
21
633
I am lacking nuance on judge objectives/rubric setup (see other comments)
31
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
30
60
9
508
31,363
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
103
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…
1
2
1
49
5,411
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
594
Overall, the obvious advice is still important. Do your due diligence. Have good benchmarks. Look at the data.
5
567
As many neolabs have transitioned to compute providers, prime has gotten worse at compute providing. makes you think
1
182
For my mental health I've decided to stop looking at the data
1
5
274
Will he or Willn't he tell us
apologizing in advance that my tweeting will likely get more vague and inscrutable for a bit. it's just that i've found some things and made some connections which spell out a clear path towards unexpected conclusions. nothing too crazy
1
8
809
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
495
who will look at those who look at data?
2
99
oh no, for i looked behind the curtain
3
91