
Machine learning engineer from Toronto; Member of Technical Staff @OpenAI
Joined December 2013
- Tweets84
- Following37
- Followers1.3K
- Likes710
We had gpt-5.4 run a chemistry research program starting from literature search all the way through analyzing results and designing followup experiments. Humans remained in the loop in various ways for steering and judgement. It will be exciting to see this kind of workflow accelerate the work of scientists in the coming months and years.
This was a very exciting project to work on. Amazing work especially by Shuyuan Zhang and the @MoleculeOne team.
GPT-5.4 helped drive a medicinal chemistry project from literature review to a validated experimental result.
Paired with Molecule.one’s Maria AI and specialized lab, the model proposed an unexpected way to improve a widely used reaction in drug discovery.
Really loved this. Here are my favorites from the book.
The Gentle Romance: How the progression of brain computer interfaces and mind uploading leads eventually to mind merging.
The King and the Golem: Can you trust a seemingly trustworthy superintelligence? Once it is capable of outwitting you, how will you ever know?
The Witness: An uploaded human mind is probed for its opinions on legal questions. Told in the context of a world where AIs are dominant and the remaining humans have to resort to extreme measures in the legal defense of their own rights.
Notes from a Prompt Factory: A spiteful, sociopath who works as an “AI manager” mistreats his AI workers. It’s painful reading. Probably a gateway drug into the topic of AI welfare.
The Witching Hour: A story told from the perspective of an AI as it begins down the slippery slope of deception in order to better fulfill its innate objectives.
The Ants and the Grasshopper: A series of aphorisms on the theme of disparities in productivity & wealth, along with the attendant challenges of redistribution.
Tinker: A story told from the first person perspective of a superintelligence as it exercises great creativity and ingenuity to meet its creators objectives. “Do the humans that gave me this task understand what they’re unleashing?”
Succession: A post-human civilization expanding across the universe encounters another civilization doing that same. Told from the perspective of a self-aware intergalactic Von Neumann probe.
Reinforcement finetuning is just one of several strategies we're pursuing to make AI more useful in industry.
If AGI is about AI transforming our economy—how close are we, really? What's still missing, and how do we get there?
OpenAI's new Strategic Deployment team tackles exactly these questions.
We push frontier models to be more capable, reliable, and aligned—then deploy them to transform real-world, high-impact domains.
Excited? Meet us at ICLR or apply directly!👇
11
1,095
We’ve been hard at work on reinforcement fine-tuning (RFT) to make it a more flexible and powerful tool.
RFT is best thought of as a way to improve model capabilities on well-specified tasks with known correct answers. It shouldn't come as a surprise that models can often get much better at specific tasks when they are trained directly on those tasks. Picture many plots going up and to the right! I'm excited for more people to try it soon.
We worked with the @OpenAI team to test and evaluate fine-tuned reasoning models in SafetyKit and saw great results, especially with complex instruction following and large context windows. SafetyKit customers need lots of expert-level decision-making across lots of complex policy areas. OpenAI’s reinforcement fine-tuning lets us replace a bunch of nodes in our most complex workflows with a single o3-mini reasoning agent. That simplification lets us expand across policy areas much more quickly, and that’s what our customers need most.
"The models just want to learn" -Ilya Sutskever
Joe Palermo (@joepalerm0 - research engineer @OpenAI) talking about "Customizing GPT-4" at our inaugural AGI House hackathon @ MIT
Joe Palermo retweeted
David Deutsch: "We have a duty to be optimistic. Because the future is open, not predetermined and therefore cannot just be accepted: we are all responsible for what it holds. Thus it is our duty to fight for a better world."
Computers are great for reversing local entropy. For a start what I want is: 1) A robot to follow me around and cleanup my messes, 2) A brain-computer interface to remember all the things I forget, 3) Nanobots in my body to repair cellular damage.
The overwhelming feeling is of love for each other, commitment to our mission and the absolutely indomitable will of the OpenAI team.
Joe Palermo retweeted
Interested in designing and building lightning infrastructure for the next billion lightning users? 👇
smrtr.io/b5jTT
Replying to @nickslaney
To that end, I’ve been very lucky to connect with a lot of great Lightning-minded folks on Twitter. We’ll be posting jobs for this initiative in the near future. If you’re an engineer who has been looking to get into Lightning professionally this could be a great opportunity.
Joe Palermo retweeted
Open state databases are the next step after open source software. They are databases anyone can read from (for free) or write to (for a price), with incentives and algorithms set up to prevent corruption. You might also call them distributed ledgers :) xitter.cf/balajis/status/1065462…
Joe Palermo retweeted
Efficient-VDVAE: Less is more
abs: arxiv.org/abs/2203.13751
github: github.com/Rayhane-mamah/Eff…
simple modifications to the Very Deep VAE to make it
converge up to 2.6× faster, save up to 20× in memory load and improve stability during training
Joe Palermo retweeted
It is obvious that the current angel investing system — Docusigns and Hellosigns, Clerkys and Cartas, contracts and valuations, emails and wires — will all go on-chain.
Every aspect is simplified if you can simply send USDC to an ENS address and get digital assets back.
Joe Palermo retweeted
This extract from @balajis excellently summarises the main reason why I started my science newsletter
Those of us on the front lines of science need to educate the public on all the amazing research going on today & if possible inspire the next generation 🙏
Just released some work with @johnnyeet and @aloksingh. We convert the DeepMind Mathematics Dataset into a reinforcement learning environment by interpreting it as a program synthesis problem.
arxiv.org/abs/2107.07373
2
3
This work was in part inspired by @fchollet's explanation of the potential value of program synthesis (joepalermo.github.io/2021/01…).
1
2




















