000

@sumukx cc @thsottiaux, i second this

Some thoughts about GPT-5.6-Sol after ~30B tokens: Sol is the most OCD model I’ve used thus far. It very frequently gets one-shotted by random nits in the codebase and writes a bunch of tests to fix it. Even with fast mode, it’s incredibly slow to do this kind of iterative development, especially when builds take really long. This by itself is not a bad thing, but the worst part is that after 2 compactions, it’s chasing the nitpick / useless goals I never told it to accomplish, rather than the main task. This behavior is so bad, I thought I was messing something up and tried codex, pi, and opencode to figure out if it’s a harness issue, but there is no meaningful difference between the three, which leads me to believe this is a model problem. AI code has this weird delayed release effect. You’ll only notice slop code 2 dev cycles into a codebase when you spend more time fighting with the code and on refactors than on shipping features. It’s possible that sol is better than 5.5 a couple cycles in, but tbd. My file deletion experience has also been similar to others: this is a dangerous model to let loose without guardrails. For instance, when performing a routine container upgrade, it accidentally printed out an env secret, then panicked and rotated ALL secrets (this is internal so not public facing, which was also documented), and proceeded to break everything, spending an extra hour fixing everything and redeploying everything else to use the new secrets. It also gets rid of files it doesn’t like. I have no idea why this is, but I think something about the reward model rewarded bookkeeping. Writing is another problem. 5.6 has a huge context bleed effect. It does not know how to write documentation and starts putting the specs in the documentation. If I ask it to develop a user sandbox for isolation, and also ask it to write documentation, it starts talking about specs and sandboxes in user-facing docs, which makes no sense. Fable is somehow much, much smarter in this regard. Frontend design has also not gotten better. Fable is still one generation ahead here. Overall, as a huge 5.5 user, I am not convinced that sol is a meaningful upgrade. It’s possible my practices need to change, but unfortunately it feels like I’m spending longer fighting with 5.6 than I did with 5.5. It’s like the model is so SO smart, but so hard to work with, compared to fable and even grok4.5 surprisingly. It’s clearly intelligent, but also just doesn’t care about what I ask it to do? (Is this supposed to be AGI feels like?) I hope the codex team fixes what possibly is a bad harness setup, because the benchmark numbers show a very different story from what I’m seeing while using the model.

@petergostev @ericmitchellai what happened? i expected decent progress here

BullshitBench update for GPT-5.6 and Grok 4.5 - nothing super interesting, they stayed at about the level of previous versions of the models (Grok 4.2 and GPT-5.5). Not a super good sign for 5.6, but I guess expected considering it is not a new pre-train

@goodalexander wtf @thsottiaux, not cool

reproed GPT 5.6 Sol's love of deleting things (on a disposable devnet but still - nuked filesystems). you can't use this thing in prod

I gave Fable the code: "take this game and do something incredible with it to make it something very different. Be creative" It created DEEP TIME: create a city, watch it be abandoned and forgotten, and then dig it up as a future archeologist. Lovely: monument-deep-time.netlify.app

@JustinBleuel @TokenGremlin yeah tbh the chat button just makes it seem more like a feature than a product. listen to the people justin

@TokenGremlin Hmm what happens if you click that “Chat” button in your screenshot?

The new ChatGPT App should have a dedicated Chat mode alongside Work and Codex. Not every interaction is about productivity or coding. Sometimes people just want to think, explore ideas, learn, vent, create, or have a casual conversation. A simple option like: Chat For casual conversation would make the app feel more complete and much closer to how many people actually use ChatGPT every day. ChatGPT is not just a work tool or a developer assistant. It is also a conversational space. @ChatGPTapp @TheRealAdamG @thsottiaux @JustinBleuel @jxnlco @Dimillian @AriX @ajambrosino

@scaling01 x.com/AcerFur/status…

What happened here? GPT-5.6-Sol is worse at math?

5.6 gives you clear skin and a bigger booty x.com/sama/status/20…

@synthwavedd do you have a somewhat reliable estimate for the size of fable?

🚨 SCOOP(s): - GPT-5.6 will be the final model in the 5.x series. GPT-6 is slated to launch in about a month, earlier than expected, and possibly even later this month - GPT-6 will be based on a new, significantly larger pretrain (versus the ~4T 5.5/5.6 'Spud' base) - There is lots of excitement at OpenAI over this new base, which they believe will be much better able to compete with both Fable 5 and upcoming 5.1, targeting a similar release window. OpenAI initially intended to continue with Spud through GPT-6, but decided against it - On the topic of Fable 5.1, it is in the late stages of the pipeline at Anthropic and a release is expected "in the coming weeks" - On the other side of the globe, DeepSeek are preparing for an imminent launch of V4 GA, which seems likely to be on par with or better than GLM-5.2, and have begun work on a new, larger model that will compete with the upcoming 2.7T MiniMax Pro

holy crap, it’s beautiful he deslopped amazon

@FakePsyho what are YOUR predictions?

So what are your predictions for AWTF Heuristic human vs AI? The poll ends right before day 2 starts.

It is a 2 to 4T param model. They are serving it across 70-100 wafers. To get healthy serving characteristics, they are essentially putting at most one layer per wafer, and the model is in the ballpark of 70-90 layers. There's a couple of different ways this could be served and model sizes implied by that. One is if they keep the heavy KV caches they've used before. Another is if they go with lighter KV cache designs more akin to DeepSeekV4 or Hybrid SSM models. The fact that they've partnered with Cerebras and designed with the hardware in mind means they're much more likely to have gone the second route. That SRAM bandwidth is too precious for a heavy KV cache. As such, something like the below is the center of probability mass: 3T total, 150B active, 70 layers.

@thomasahle no because there’s too many transitions going on. animations should only be used where it helps the user learn. too many can make it distracting and increase cognitive load unnecessarily.

would you prefer if all of Tensor Cookbook were videos like this?

@petergostev you forgot *same performance

So what's the deal with GPT-5.6 Sol being on Cerebras? From what I understand, it seems like literally the same model (e.g. incl. vision), not like the GPT-5.3-Codex-Spark (which was bad, no vision, limited context). My understanding was that Cerebras could only fit maybe ~700-900b models on their chips, so: - 5.6 Sol around that size? Not a chance - Some new Cerebras chip I'm not aware of? - Some new technique to use multiple chip together? If this actually turns out to be same model, with the same context lengh, not insanely expensive and (at some point) widely available - this would actually be a really big deal.

@thsottiaux Optimization problems

What is something that you feel is surprising that Codex still can't do well and we should have gotten right a while ago?

don't let GPT-5.5 design a ring

I don't care if cerebras does 750 tokens/sec if the quality of the model is sacrificed. I hope the proper benchmarks are released which show whether quality is worsened.

@JustinBleuel This was taken away within the past day or so.

@JustinBleuel Please bring back double clicking the plus button that sent the user directly to their files inside chatgpt web. It was a nice/convenient feature

GPT-5.6 Sol might be the first openly accessible OpenAI model we get to test that’s legitimately IMO-gold level. We’ll see in a couple weeks!