matt hardy
@mdahardy
Followers
1K
Following
3K
Media
122
Statuses
465
cto @roundtablehq_, prev phd @princeton // language models, cogsci, ml
san francisco
Joined April 2014
This is really cool to see! Amazing visualisations too ;) it'll be interesting to apply the techniques we used to understand human decision making in this game to the LLM reasoning process too. Cc @weijima01 @cocosci_lab @marcelomattar are y'all interested?
New LLM game tournament 🎲🕹️! Four in a row from @basvanopheusden and @Ikuperwajs's "Expertise increases planning depth in human gameplay" ... and our reasoning depth analysis. 🧵
1
4
9
Personal AI as a guardian angel is a very cool idea
IMO the ultimate interface for AIs is no interface. Not chatbots, not prompting. @goldenkrishna has been on this for years. Last decade it was screens, now it is chat interfaces. Tesla solved the car unlocking problem outlined below (just walk up to the car with your phone in
0
0
3
If it’s not clear from the context, I never use non-reasoning models. I don’t have time and patience for that
Am I the only one wondering how come the models hallucinate so little in 2025?
12
7
291
We are just getting started
Starting a new fund called: Athletic Gingers We believe they are the key to new technology Elite genetics for unique thinking & incredible stamina for company building Huge upside guaranteed
2
0
10
It's hilarious how "bi-weekly" is both precise in its two meanings and yet never clear from context which meaning is being used. We should get better terms.
3
0
4
We're on the front page of HN!
1/ Come join us on Hacker News today. What happens when you benchmark different AI Agents (GPT, Claude, Gemini) on the CAPTCHA?
0
2
7
We ran the latest AI agents against Google's reCAPTCHA v2. Claude 4.5, Gemini 2.5 Pro, GPT-5. Short thread on what we found. 🧵
1
3
4
A @cocosci_lab member building robotic vision systems. Really cool company from @gianlucabencomo!
We already have the best models @EfferenceAI And we’ll be shipping the best cameras for robotics in March: https://t.co/EhgRP2LFOI Excited to announce @EfferenceAI and my participation in @ycombinator
1
0
3
what's the best model for cursor? asking for someone who is a year late and just trying it now.
2
0
2
This is why I'm so excited about proof of human. If you don't know who's human and who's ai, the incentive to interact or trust anything online plummets
What happens when online job applicants start using LLMs? It ain't good. 1. Pre-LLM, cover letter quality predicts your work quality, and a good cover gets you a job 2. LLMs wipe out the signal, and employer demand falls 3. Model suggests high ability workers lose the most 1/n
0
2
9
The biggest lesson from YC (@ycombinator) isn't about monetization or network effects. It's about velocity. Here's the failure mode: deeptech founders mistake research progress for startup progress. They spend 18 months building "the right foundation". It feels virtuous. It's
1
1
8
"Why can't an AI learn to fake human behavior?" It’s a fair question. AI agents optimize for a global, optimal solution. Humans do not. Our function is messy and filled with hesitation and doubt. In this video, I walk through two simple demos from our research site: ---
0
3
7
this can't be the story because the slide is still open right?
never knew the full context behind the Boston Slide Massacre and this makes the clip even more satisfying now
0
0
2
crazy that "vibe coding" (the phrase) has only existed for 8 months
0
0
5
You can just prompt things! And we want to make it easier for everyone to use these methods. As such, we're very excited for @ExpectedParrot to be part of the Fall YC Batch!
You can just prompt things
30
25
287
A funny thing with Google search AI mode is that it always does a web search for every response. So when I say "Thanks!", it searches for Thank you images it can respond with
0
0
3
What happens when a system built to protect humans can’t tell what a human is anymore? Concert fans are getting blocked by ticketing bots. Grandmothers are struggling with CAPTCHAs. You’re drowning in AI slop. We may not love clicking traffic lights, but fraud is real. The
0
1
8