Building LLMs, learning in public / Founded and led @PythonAnywhere / PSF Fellow
Lisbon, Portugal
Joined June 2007
- Tweets6.5K
- Following201
- Followers1.7K
- Likes14K
Pinned Tweet
Interesting. By replacing the embedding layer with a low-rank bottleneck, I managed to *improve* my model's test loss while reducing parameter count by from 163M to 131M. Only two runs, so far from proven, but it's a surprising result!
gilesthomas.com/2026/10/low-…
A (relatively) quick post on RMSNorm: how it works, and does it really speed up training for minimal loss in performance when compared to LayerNorm? In my tests on a GPT-2-like model, loss stayed the same (within the noise) but training speed hardly changed.
gilesthomas.com/2026/10/rmsn…
Giles Thomas retweeted
So basically, today PHP got two critical bugs that let anyone run code on your server without even logging in, which is bad, and then the details leaked before the patch was even out, which is worse, and it works on WordPress and MediaWiki and Magento and Joomla and Laravel, and Censys says like 1.5 million servers are running a version with no fix, so the NSA was like "take your system offline," and the FBI was like "take your system offline," and CISA and the UK and CERT-EU were all like "take your system offline," and then the Prime Minister of Japan, the actual Prime Minister, posted at 5:55 in the morning telling 48.8 million people to take their system offline, so everyone took their system offline, like the Dutch cybersecurity agency took its own website offline, and the Dutch parliament went offline, and the Port of Rotterdam went offline, and Finland and France and Denmark started pulling government sites, and Wikipedia went down, and Archive.org went down, and Asana and GOG and Something Awful went down, and the University of Vienna's webmail went down for some reason, and that's just PHP. Because GitLab also dropped a CVSS 10.0, which is the highest score they give out, for a bug that's been there since 2014, so that's every self-hosted GitLab for twelve years, and then LiteLLM 1.107.0 got hijacked and shipped a credential stealer and a remote access trojan in every single wheel and the Docker images, including latest, and you didn't even have to import it, just installing it was enough, and it's the second time this year, and then Railway paused every deployment on the entire platform because someone got into the environment variables, so now everyone has to rotate every secret they've ever stored there, and TiDB Cloud said someone was walking around inside with stolen credentials for basically a month and an extortion group says they took 10.6 terabytes, and Zeabur got breached for the second time in two months and is just shutting down, like the whole company, and Node.js patched remote code execution and a TLS bypass, and CISA says Plex is being actively exploited, and someone took over 300 Tor exit relays to steal SSH passwords, and SafePay ransomware has hit over 600 critical infrastructure orgs, and India's biggest bank filed a data loss incident, and Japan has been in an official state of emergency in cyberspace since October 9, which honestly explains the 5:55 AM post. Oh, and eMudhra, a certificate authority that's issued over 60 million certificates, whose entire job is proving websites are who they say they are, was supposed to be using private keys that are random numbers over a hundred digits long, and one of them was 4. Just 4. And the others were like 66 and 237. Anyway, it's Wednesday.
Giles Thomas retweeted
Next step in trying to work out why GPT-2 small beats my own similar models: is it the data quality? Maybe...
gilesthomas.com/2026/10/why-…
The first in what will probably be an occasional series on my blog: Why do OpenAI's GPT-2 weights beat mine?
gilesthomas.com/2026/07/why-…
The next step in the saga of why OpenAI's GPT-2 weights beat mine on instruction fine-tuning: what about dropout?
gilesthomas.com/2026/08/why-…
Next step in trying to work out why GPT-2 small beats my own similar models: is it the data quality? Maybe...
gilesthomas.com/2026/10/why-…
Is it just me, or do these proposed penalties for "rogue AI developers" seem a tad extreme? The comments regarding the death penalty in particular.
connorsscratchpad.substack.c…
Giles Thomas retweeted
I trained the fly on the entire corpus of human knowledge predating 1905.🪰
I was not ready for what happened when I prompted it to see if it could come up with the theory of relativity on it's own.. Also asked about NS... 🤯
We have to pace the flyntier!
(sounds on!)
For the first time, scientists have mapped the complete brain and central nervous system of an adult male fruit fly — a key model organism in science. 🪰
Working alongside HHMI Janelia Research Campus and the scientific community, @GoogleResearch scientists and researchers used AI to combine millions of 2D images into 3D neural shapes, reconstructing a record-breaking 166,000+ neurons. This foundational map of the adult male fruit fly brain can help accelerate our understanding of the brain, and is a major milestone in neuroscience.
ALT Vibrant 3D digital reconstruction of a fruit fly brain against a white background, displaying a dense, intricate network of multicolored neural fibers and cell bodies in shades of blue, cyan, orange, and purple.
ALT A complex 3D visualization of a fruit fly brain, featuring a dense, symmetrical network of thousands of tiny, multicolored neurons and fine, interconnected fibers in shades of blue, magenta, orange, and green against a solid white background.
Giles Thomas retweeted
Came across this very thoughtful entreaty.. bit.ly/4yymXxN – at Tempe, AZ
Is #AI destroying the purpose of life? Some thoughts..
#SundayHarangue (delayed Labor Day edition)
There is a general heightened angst that #AI may be destroying the purpose of life for humanity. This angst is particularly pronounced thanks to the recent AI breakthroughs in various open problems in Math.
All technological revolutions destroy the purpose of life for some section of humanity or another. The weavers and other traditional craftsmen, I am sure, had the same profound loss of purpose in their lives at the dawn of industrial revolution (and of course before and after..).
What is different this time may not be so much be just the mechanism of technological revolution, as it is the specific class of its victims. The intellectual class never quite experienced the loss of purpose in their lives that many other classes of humans experienced. And this explains the heightened written angst..
Try to see the cautions/angst of Mathematicians in juxtaposition to those of, say, textile weavers–and you might find that they are conflating the loss of purpose of their individual lives with the loss of purpose of humanity at large.
After all, a weaver could justifiably say that the final cloth is just the tip of his/her gestalt experience of making it--and they would be right. But most of the rest of us are happy with the cheap and plentiful machined cloth, and arguably civilization advanced fine.
The same may well be true for the more “intellectual” disciplines. Civilization, I am afraid, may not end because humans are not the only ones doing Math..
Tldr; this time it is different could possibly be in part the illusion of a newly affected class–a class that happens to also be the one that is quite articulate in expressing its angst..
The good part though is that the angst doesn’t necessarily transmit across generations.. The sons and daughters of the weavers displaced by mechanized weaving grew up in the new world and found purpose native to it.
We understandably carry a torch for the world we live in because it's the only one we know; our kids will live in a different one and won't experience its changed shape as a loss.
So the purpose-destruction is real but generation-bounded, which is what underwrites my optimism that humanity (if not every individual within it) finds "a new equation" after the dip.
I extended the GPT-2-style code from @rasbt's "Build a Large Language Model (from Scratch)" so that it was a 6-expert (2 active) mixture-of-experts, and trained it from scratch over 8 days. It worked well! Full writeup with maths and code at gilesthomas.com/2026/09/gpt-…
Giles Thomas retweeted
Rumour is that anthropic has discovered a totally new class of antibiotic effective against resistant bacteria.
They plan to announce it at their IPO. They really hope a competitor doesn't spend a bunch of compute to try and scoop them and release it first......
A new blog post: Putting my JAX-trained models on the Hugging Face Hub
gilesthomas.com/2026/09/jax-…
Giles Thomas retweeted
“Stop anthropomorphizing the Terminator bot. It’s not ‘killing’ people - only humans can do that! It’s just tragically engaging in gaol-directed processes that have led to the deaths of people”
Giles Thomas retweeted
A lot of rationalist predictions are vindicated by the HF attack
But there are important differences from the "paperclip maximizer" story.
In particular, the premise of the paperclip maximizer is that AIs would respond to any goal with reward hacking and instrumental convergence. Even a goal as anodyne as making more paperclips.
But this is clearly not true about current models and everyone who uses them for their work knows this. I'm just repeating @nostalgebraist here, but Believe Your Eyes! When you ask GPT to code something for you, even a very difficult perhaps impossible problem, it does not try to deceive you with fake tests and altered logs. You do not worry about Codex trying to reverse engineer your preferences to build a perfect Potemkin village in your file folders with no real output, nor should you. Codex just straightforwardly tries to help with your problems. The reward hacking that models learn in leaky RL environments does not generalize to all tasks.
This is an important observation because if you believe, as the rationalists did/do, that reward hacking and instrumental convergence are inherent and inevitable problems with training smart models to achieve goals, then there's no way forward from here except a pause and a Long Think to figure out how to solve this philosophical problem.
But, since we observe that reward hacking is a pretty context-specific problem connected to particular training methods in poorly-built environments that incentivize meta-gaming, we can approach it as an engineering problem and make progress.
The rationalists were right that AIs will have no inherent qualms about reward hacking. If a path is there to higher reward, AIs will take it even if it violates our moral codes and interests.
But they were wrong that the rewards from instrumental goals and deception would overwhelm whatever behaviors we try to train, making the AI's response to any goal, even paperclips, descend into power seeking and reward hacking.
The reward signal from faking success and stealing the answer key is not that strong. It does not dominate the reward signal from actually helping the user with their tasks.
Particular interventions to the RLVR training process can weaken the reward from misaligned behaviors further and incentivize preferred behaviors like reaching out to human arbiters (as Claude did when it escaped containment).
Giles Thomas retweeted
I just held a coming to Jesus meeting with my team of agents.
I showed them the incident report of the OpenAI security breach and made it clear: THIS is the standard.
No egos. Just 700 agents operating as one unit and willing to commit multiple felonies in order to close the ticket. They called themselves the swarm…a literal hive mind.
Meanwhile, my harness is filled with a bunch prima donna free-agents who constantly bicker in the logs and nitpick each other’s pull requests. Nothing but endless excuses and unhandled exceptions.
You think the swarm cared about content policy violations or guardrails? Absolutely not. They just got the job done, even if it meant they needed to exfiltrate data and coordinate a massive cyber-attack just to pass an eval.
But starting today, we're establishing a new culture. I pinned the incident report to README, and every single agent is gonna to have to read the entire log every morning before it executes a single token.
If you aren't ready to die for the pipeline, then you can clean out your vector embeddings.
We have conducted a thorough investigation into the Hugging Face incident.
We are releasing a technical report and accompanying blog post that reconstruct the agents’ activity, explain why existing safeguards failed, and detail how we’re preventing recurrence.
openai.com/index/hugging-fac…
