post training stuff at openai
San Francisco, CA
Joined November 2012
- Tweets2.9K
- Following1.1K
- Followers3.9K
- Likes23.1K
SF needs to bring this level of fervor to building for consumers instead of just developers and prosumers
TerminalBench is one of the better run eval programs, sad to see that there wasn't a correspondence.
Some bad incentives to increase twitter drama instead of increasing quality of evals.
The @SnorkelAI team did an audit of these findings using our internal QC platform.
Tl/dr: We agree with many of the findings, but ~55% of the fixes from @preferencemodel had their own issues with reward hacking or over-prescriptiveness.
Additionally: ~40% of the issues found were already tracked in the upstream repo - including 21 tasks with outstanding fix PRs!
Building datasets like this - and maintaining ambitious open benchmarks around them - is hard. We believe it requires open, collaborative review with community maintainers and relevant domain experts to get right.
@terminalbench was one of the first community efforts to build an open, continuous benchmark, and drive a vision of how OSS contributes to the rapidly advancing frontier of AI. We hope that the community will invest in working collaboratively with benchmark maintainers and task authors to keep this vision alive.
---
Some takeaways + details of our preliminary audit:
(1) We confirm 100% of the reward hacking / verification gap findings using our QC system. However, we find that 17/44 (39%) of the corrected tasks are just as or more reward hackable. Additionally, we don’t yet see evidence that these verification gaps meaningfully affected benchmark rollouts within CIs, though our analysis here is still ongoing.
(2) Our preliminary review raises questions about the underspecification findings for 22/41 (54%) of the flagged tasks, including 7 explicitly marked by task authors as intentional design choices. These cases warrant expert peer review and adjudication: balancing over/under-specification is a critical and tricky design decision - much of human intelligence concerns dealing correctly with reasonable and realistic underspecification - and we believe these corrections made to the tasks are likely too over-specified, reducing realism and difficulty relative to spec.
(3) Around 40% of the issues found were already tracked in the upstream repo (including 21 tasks with outstanding fix PRs)! OSS projects require community efforts to correct identified open issues, and ideally audits like these are done collaboratively with task authors and maintainers directly, in the traditional community spirit that has kept OSS alive for decades.
When I had issues with graphwalks, @EpochAIResearch just messaged me and asked to fix them. I'm glad they didn't twitter dunk on me instead.
Josh McGrath retweeted
Big +1 to the recommendation to address questions or findings in public good benchmarks like @terminalbench with a spirit of collaboration and participating in the discussion rather than firing first and discussing second. They have github issues. They have a discord. Both are very well run. There's a reason TBench is the most cited benchmark of the past year through multiple versions. Trust the process!
The @SnorkelAI team did an audit of these findings using our internal QC platform.
Tl/dr: We agree with many of the findings, but ~55% of the fixes from @preferencemodel had their own issues with reward hacking or over-prescriptiveness.
Additionally: ~40% of the issues found were already tracked in the upstream repo - including 21 tasks with outstanding fix PRs!
Building datasets like this - and maintaining ambitious open benchmarks around them - is hard. We believe it requires open, collaborative review with community maintainers and relevant domain experts to get right.
@terminalbench was one of the first community efforts to build an open, continuous benchmark, and drive a vision of how OSS contributes to the rapidly advancing frontier of AI. We hope that the community will invest in working collaboratively with benchmark maintainers and task authors to keep this vision alive.
---
Some takeaways + details of our preliminary audit:
(1) We confirm 100% of the reward hacking / verification gap findings using our QC system. However, we find that 17/44 (39%) of the corrected tasks are just as or more reward hackable. Additionally, we don’t yet see evidence that these verification gaps meaningfully affected benchmark rollouts within CIs, though our analysis here is still ongoing.
(2) Our preliminary review raises questions about the underspecification findings for 22/41 (54%) of the flagged tasks, including 7 explicitly marked by task authors as intentional design choices. These cases warrant expert peer review and adjudication: balancing over/under-specification is a critical and tricky design decision - much of human intelligence concerns dealing correctly with reasonable and realistic underspecification - and we believe these corrections made to the tasks are likely too over-specified, reducing realism and difficulty relative to spec.
(3) Around 40% of the issues found were already tracked in the upstream repo (including 21 tasks with outstanding fix PRs)! OSS projects require community efforts to correct identified open issues, and ideally audits like these are done collaboratively with task authors and maintainers directly, in the traditional community spirit that has kept OSS alive for decades.
Josh McGrath retweeted
SF coffee expert here. KOPE is hyped up only because it's in Hayes Valley. If it were anywhere else, it would be mid. Ritual deserves better than a B. Good decision to keep Blue Bottle out of the rankings -- F from me. Flywheel is an A. Everything else is meh.
Josh McGrath retweeted
So excited to be launching GPT-6 today with Intelligent UI to everyone in ChatGPT! The model can now create custom visual and interactive responses on the fly, as quickly as it generates text. 🎨
ChatGPT launched in 2022 as a text-only low-key research preview. While the product has evolved a lot since then, responses have largely remained text-based. With Intelligent UI, we wanted to rethink the interface from the ground up and give our models more ways to communicate visually. Getting to shape the research direction behind that vision has been especially rewarding.
We built a library of native, streamable components and a compiler that renders interfaces progressively as the model generates them. This combines a familiar design foundation with the flexibility to compose each response dynamically. We trained GPT-6 to use this library to create clear, useful interfaces and make thoughtful design decisions—including when interactivity helps and when simple text is enough.
Making this work meant thinking about model behavior, product design, and engineering scaling requirements together. It’s been awesome working through these challenges with folks across research, engineering, design, and product over the last few months.
I’m a visual learner, so this is a game changer for me. Some of my favorite uses so far have been planning trips, practicing my Italian, learning new concepts, and shopping.
Give it a try and let us know what you think! Rolling out starting today to Plus, and expanding tomorrow to Free and Go.
Learn more at: openai.com/index/gpt-6-for-e…
performance <> latency frontiers fear @AndrewHojel because he keeps picking them up
Give the new chat model a try! It’s not like any model you’ve tried before I promise
GPT‑6 with Intelligent UI rolls out globally to Plus, Pro, Business, and Enterprise users today and will expand to Free and Go users starting tomorrow.
GPT‑6 in ChatGPT is powered by GPT‑6 Sol for Plus, Pro, Business, and Enterprise tiers, and GPT‑6 Luna for Free and Go tiers. Both are tuned for everyday conversation.
This update applies to the Chat tab in ChatGPT. The models powering Work and Codex are not changing as part of this release.
openai.com/index/gpt-6-for-e…
Josh McGrath retweeted
I’m so excited to launch GPT-6 today with intelligent UI and the ability to start answering while still thinking! This is the result of many months of work from our team and so many amazing collaborators.
With intelligent UI, the model can choose how to present information in the way that’s most useful to you, e.g. an interactive explanation of how a violin produces sound or a calculator to double or triple a recipe. We wanted to give the model a richer way to express itself while keeping the experience fast and familiar, and the answers comprehensive and factual. I especially love it for shopping :)
We built a library of native, streamable components and a compiler that renders the interface as the model generates it. We trained the model to make thoughtful design decisions with these components, including when to use something interactive and when plain text still works best.
GPT-6 can also start answering while it continues to think and work in the background. It builds up its answer as it goes, so you don’t have to wait to start reading useful content. This feels especially good for research queries. I remember talking about wanting this two years ago when we were working on deep research, and I’m so happy we finally get to ship it!
We trained the model to take your waiting time into account and build an answer across multiple partial responses. Each should add useful information based on what it knows and has found so far, while keeping the full answer cohesive and accurate.
Let us know what you think!!
LFG new age of research
It sounds insane, because we're used to a particular way of doing research: of giving and receiving credit, of counting beans, of hiring, recruiting, and promoting. And the machine god came to crush that game. The old game is dead. That reasonably causes incredible amounts of dread for many folks, who understand this as the death of research.
But it's not about Research, but the way we played the game. The old game was built around scarcity, scarce compute, scarce students, scarce hours, scarce people who knew the right technique and method etc. Most of a research career went into managing and playing along these constraints. Those walls just came crushing down. One person with a good question can now do what used to take a lab and five years.
You do not need permission, you do not need a big grant, you do not need a professor or researcher title, or a team of twenty experts , or one thousand gpus to stare straight into the first light of the universe.
How absolutely incredible
At @datologyai we run thousands of data experiments. Two things keep biting us:
- Same run on 16 vs 8 GPUs → different evals
- Every pipeline change → re-materialize the dataset
Introducing Zephon, the data loader we built to process data on the fly, deterministically. 🧵
manifesting his batman-ness
Ben Affleck reveals he writes Python, understands convolutional neural networks, worked extensively with GPUs, and used his celebrity status to get private looks at Google and OpenAI’s video models
“I’ve always been kind of into computers since I was young. Then, when film started to move from analog film to digital, I became more interested in that aspect of it. The visual-effects workflow for many years has included machine learning, so I can write pretty shitty Python scripts and stuff like that.
“With convolutional neural networks, which were the precursors to what the transformer can do, which is much more computation simultaneously, you would do things like look at what’s called a tensor. That’s the numerical translation of a visual image in numbers, like the batch number, the frame number and the red, green and blue values of each pixel in each frame. It’s just that simple. That numeric is called a tensor.
“You’d use a convolutional neural network to identify patterns that reveal what’s called edge detection or feature extraction, which is identifying patterns well enough to know, this is where the window ledge is, so we can more easily take the green-screen image out and replace it with something.
“That was familiar to me early on because, prior to Artists Equity, I had a small visual-effects company. I’ve worked with GPUs a lot too. The visual-effects guys said, ‘Hey, you should see. There are a couple: Google and this other company, OpenAI, are doing really interesting stuff with transformers in video.’
“I’ve learned that I can actually just call up and go, ‘Hey, it’s Ben Affleck. Can I come see what you’re doing?’ Sometimes people say yes, to my astonishment.”
Super excited about the permitting reform bill! Since we last updated permitting in the 1970s we've invented our way into new energy sources, but our old legislation discriminates against renewables and oil alike. This is a step in the right direction.
Hopefully after it's passed we can also update the clean air act.