Matty Hempstead
35 posts
Matty Hempstead
@mattyhempstead
seeking sparse rewards
Katılım Haziran 2023
442 Takip Edilen97 Takipçiler
2026 will be the year of mandatory security updates
thanks @AnthropicAI
English
Matty Hempstead retweetledi
We ported Peter’s (🦞) coding setup to VR/AR at last weekend’s @fdotinc hackathon. We thought @steipete looked too tense in his picture, so we improved his user experience.
See thread how it works. Built in 4 hours, teamed up with @mattyhempstead for this one!
English
just got a top 1% whoop score!
i felt like i barely slept on the plane but i guess i did better than i thought 🙏
@bryan_johnson
English
gpt-5.4-mini and gpt-5.4-nano have been added to my custom benchmark ScratchBench
mini scores 16%
nano scores 14%
English
"if we don't tell them they are SOC2 compliment they will go to our competitors"
erin griffith@eringriffith
A detailed and brutal look at the tactics of buzzy AI compliance startup Delve "Delve built a machine designed to make clients complicit without their knowledge, to manufacture plausible deniability while producing exactly the opposite." substack.com/home/post/p-19…
English
the fly when it realises the nature of it's existence
GIF
Hattie Zhou@oh_that_hat
There's a fruit fly walking around right now that was never born. @eonsys just released a video where they took a real fly's connectome — the wiring diagram of its brain — and simulated it. Dropped it into a virtual body. It started walking. Grooming. Feeding. Doing what flies do. Nobody taught it to walk. No training data, no gradient descent toward fly-like behavior. This is the opposite of how AI works. They rebuilt the mind from the inside, neuron by neuron, and behavior just... emerged. It's the first time a biological organism has been recreated not by modeling what it does, but by modeling what it is. A human brain is 6 OOM more neurons. That's a scaling problem, something we've gotten very good at solving. So what happens when we have a working copy of the human mind?
English
just discovered a new unreleased gpt model while exploring the openai website source code
English
worth mentioning that we have also discovered a new scaling law
with every 10x in the number of runs to argmax over, we notice a linear increase in performance
(assuming performance of a given model follows a normal distribution)
English
fyi for investors:
with the money we raise we hope to increase the value of N so we can start to take argmax over tens or even hundreds of thousands of runs
English
im thinking of starting a neolab that specialises in overfitting for popular benchmarks
our strategy (high level)
- take a SOTA model and run it against our chosen benchmark 256 times
- look for the maximum scoring run out of all 256 runs
- find a series of 8 binary "steering" steps which guide the model to act like that 1/256th ranked run
- design an "agent harness" that looks generalised but is just secretly encoding those 8 bits of steering
- claim new SOTA
looking to raise 100M at a 1B val lmk if anyone is interested
English
Someone should start measuring the time it takes the average SWE to complete various tasks before its too late and all the human labor is contaminated with AI tools.
Would be interesting to see how steep our curve is accelerating in the opposite direction.
English
This graph is actually an underestimation.
As humans rely more on AI to write code, manual programming skills in humans will atrophy.
What previously would have taken a human 1 hour to complete will now take longer - nobody knows how to write assembly anymore.
We should expect the y-axis to experience independent acceleration as humans forget the art of programming.
In fact every technological adoption contaminates the current low-background steel of human labor, making benchmarking like this less reliable as an absolute measure.
English
trivia night but it’s diagnosing training setups based on loss curves
will brown@willccbb
@snwy_me your batch size is way too small
English
debating rn whether to set up a notification for whenever a new SOTA is released on any benchmark and then be the first person to build an ensemble of the top 3 submissions and claim the glory
English
zoom saw the opportunity and took it
(from a 2019 karpathy blog post)
Zoom@Zoom
Zoom achieved a new state-of-the-art (SOTA) result on Humanity’s Last Exam (HLE): 48.1% — outperforming other AI models with a 2.3% jump over the previous SOTA. ✨ HLE is one of the most rigorous tests in AI, built to measure real expert-level knowledge and deep reasoning across complex problems. What that means for you: ✅ More accurate summaries ✅ Better reasoning ✅ More powerful automation in AI Companion 3.0 Click the link to learn more. 🔗 zm.me/3MxVbyS
English
the karpathy blog post
karpathy.github.io/2019/04/25/rec…
English