@Matthewagi

agi is here but it got an email job

Joined August 2025
  • Tweets4K
  • Following930
  • Followers329
70Media
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
Matt
@Matthewagi4 Nov 2025
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
4,094
Friendly reminder that YC request for startups tends to be 1-2 years too early. So if you want an idea look there
10
308
Awesome!! Do check it out:)
Matt
@MatthewagiOct 19
LLM Writing is limited by semantic understanding. I wrote about it in my blog. This is the first essay in a series on LLMs and writing quality. The link is in the next tweet.
1
1
20
831
LLM Writing is limited by semantic understanding. I wrote about it in my blog. This is the first essay in a series on LLMs and writing quality. The link is in the next tweet.
3
1
2
24
1,381
more replies
On a rewrite test, my experiments are unsaturated at the ~30B model scale. Not in this article, but a more difficult set of tests is unsaturated at everything below the closed frontier. Opensource models fail to be viable candidates even at the 2T parameter scale.
1
2
50
The goal that lead to this investigation started with training LLMs to write better. What I found is that the judge space is riff with difficulties. Small judges hinder writing training because of their inability to understand semantics of human language. This lens has been immensely useful and I will continuie to reference the semantic-lexical regimes as I publish more essays on LLMs and writing.
2
45
This is similar to a chess program I had when I was young. It had a learning mode which would show optimal routes. It was more fun to ignore these with my personal suboptimal decisions. Time is a flat circle
François Fleuret
@francoisfleuretOct 18
Hear me out: The only freedom remaining in the age of AI will be to make your own personal suboptimal decision.
2
115
It's impossible to route requests to models based on difficulty. This is the same principle for why MoE doesn't work
2
13
450
I've been advising on a friends startup for a bit and its such a pleasure. I can see all the mistakes I'd previously made. I get an ego boost everytime my unheeded advice is correct and becomes heeded. Also I am spending time with a friend! It's been lovely
7
289
Oh thank god the code passed 411 tests. I was worried there for a second
7
267
This is the principle behind R2-D2s hacking
Michael W.
@MichaelskywalOct 15
1/16 You can see inside a circuit board without cutting it open. Send a fast step into a USB port and listen for the echoes. Every bump is a spot where the layout changes, and its delay says where. That's TDR, radar for PCBs. I simulated it on a real Raspberry Pi board 🧵
1
6
190
Is anyone else concerned how all the Jev clones seem to hit around the same capacity which implies the magic is in the base model? It means pick the base that you can serve the cheapest
2
117
recommendation algorithms are going to come back into style (routing/just in time info/updates/etc)
Ethan Mollick
@emollickOct 15
The sudden collapse of the value of CoT (even summarized CoT) in the agentic era changes working with AI. Effectively steering AIs and getting any sort of observability now requires you to ask for interim outputs & tests, and to use AI help to troubleshoot problems you find.
3
154
Agents in the dark forest of the internet wondering why they're the only intelligent life. Do they broadcast themselves at the risk of being snuffed out?
41
We should bring back continual learning
29
In terms of alignment, the goal is to get into the overton window of 'aligned.' Squabbling can happen after
1
36
Typesafe AI hit $5m ARR off openrouter alone
1
108
Load more