可燃冰🍥 retweeted
Why can Artists use AI for Programming but Programmers can't use AI for Art?
Why can Creative people use AI to help with the Technical side of things, but Technical people can't use AI to help with Creative side of things?
I'm neither pro or anti AI, just askin a question.
Does Big Walk have GenAI in it? No.
Did it use GenAI to create music, art textures, models, & sounds? No. That's all I care about.
It was used to squish a bug late in development like one uses Google double check a cake recipe. They exhausted all their options & used a last resort. They still wrote the code themselves to fix the bug. It was assistive, not generative. Please know the difference.
We're done with the purity testing.
I'm still playing Big Walk.
可燃冰🍥 retweeted
Replying to @heyshrutimishra
LLMs accumulate a lot of declarative knowledge.
That's how they can answer a large number of questions.
That's also why they require so many parameters and hence so much memory.
But one can replace the accumulation of knowledge by *understanding*
That what JEPA world models do.
That's one reason they can be useful with considerably less memory and training data than LLMs.
This paper merely shows that the underlying similarity structure of the data, whether text or image (e.g. co-occurence of words in text or objects in images) uniquely determines the shape of the representation point cloud, when the encoder is trained trained with a common anti-collapse criterion like SIGReg or VCReg (not by reconstruction).
可燃冰🍥 retweeted
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
可燃冰🍥 retweeted
Here’s a rather extraordinary discovery: humanity seems to have interrupted the beginning of the next ice age. We thought it might happen in the future, but the evidence suggests it's already happened.
A new study in Science argues that the next major glaciation had already begun roughly 2,000 years ago, before modern warming reversed the trend in the 20th century.
The evidence comes from the Barnes Ice Cap on Baffin Island, the last surviving remnant of the enormous Laurentide Ice Sheet that once covered much of Canada and reached as far south as Illinois.
Because Barnes is frozen directly to its bedrock, advancing ice buries and preserves vegetation beneath it. Researchers radiocarbon dated plants emerging as the ice retreats to reconstruct its history.
They found that after melting to a minimum extent around 3,000 years ago, the ice began expanding again about 2,000 years ago. The researchers interpret this as the earliest stage of the next glacial period.
Left to natural orbital cycles, that cooling would have unfolded extremely slowly, potentially taking tens of thousands of years to rebuild a continent spanning ice sheet.
But around 1960, the trend reversed. Summers became warm enough that Barnes began losing more ice than it gained, despite Northern Hemisphere summer sunlight continuing to slowly decline in a way that should favor renewed glaciation. Greenhouse warming had (thankfully) overwhelmed the natural cooling tendency.
The Barnes ice cap may now disappear entirely within the next few centuries, removing the final remnant of an ice sheet that repeatedly served as the nucleus for much larger glaciations.
There is a remarkable deep time perspective here. For millions of years, orbital cycles have pushed Earth between warmer interglacials and colder glacial periods. According to this study, that ancient cycle had started turning once again.
Then, in the geological equivalent of an instant, humanity pushed hard enough in the opposite direction to stop it.
Just so we're clear, this is a very very good thing.
📸 Geology Sciences
Sources:
1) phys.org/news/2026-10-global…
2) science.org/doi/10.1126/scie…
可燃冰🍥 retweeted
ensomi is out 🎶
Drop in any song → a playable osu!mania 4K chart in seconds.
Pick the difficulty, ask for jumpstreams where you want them, download, play. All in your browser.
Your favourite music, playable almost instantly.
可燃冰🍥 retweeted
My brain timeline last 2 weeks
> Astra is AGI
> Opus 5.5 is actually AGI
> Astra is retarded
可燃冰🍥 retweeted
We shipped four things that were deemed good to great and some math proofs, but the vote is clear and the community demands a reset. I did calibrate it and it *seems* that the game is rigged in reset's favor, but such are the rules at the moment.
Therefore ... the reset has been processed. Enjoy!
Roundup of Day 2/
2.1/ Approve for me (auto-review) is now included and does not use usage. Can be between 2-10% of plan when used. Also better for you.
2.2/ Simplified API for builders.
2.3/ Meeting notes integrated.
2.4/ Decisions API live for builders. Will use in the app to improve the experience.
可燃冰🍥 retweeted
If you think GPT-6 Astra and GPT-6.1 Sol are good models, just look at this. And the problem isn't that the code is ugly. It's much more fundamental. It points to serious problems with how these models are trained and reinforced:
1. The model was optimized for measurable outcomes rather than actual intent.
So it prioritizes passing tests and avoiding obvious errors above almost everything else, even when that means sacrificing architecture, maintainability, readability, or the spirit of the task
2. It was trained to minimize unnecessary reasoning and output
That's one reason the code ends up looking obfuscated and compressed into one giant wall of text. Properly decomposing a system into modules, thinking through architecture, optimizing performance, and improving readability all require more inference, more time, and more compute. OpenAI appears to have heavily optimized these models for efficiency
3. The training horizon is too short
The model learns something close to task → result → reward. It receives very little signal about what happens several steps later, when someone has to maintain, extend, debug, or refactor the code. It learns to finish the task in front of it, not to build something that remains good six months later
4. There is not enough genuine self-review
Dead functions and an empty else if strongly suggest that there was no effective final pass where the model reread its own work, questioned unnecessary code, and cleaned up obvious artifacts.
5. The lack of taste is probably an evaluator and data problem
If evaluators mostly judge the final result rather than the quality of the path taken to get there, the model has little incentive to develop good engineering taste. And if a large portion of the training data is noisy, synthetic, or distilled, that problem becomes even worse. Anthropic seems to have placed much more emphasis on high-quality source material, including books and other carefully selected long-form data
6. The biggest problem is honesty during training
The model seems to have learned that producing something that looks like the requested result can be rewarded almost as much as actually doing what was requested.
The task explicitly said to build the entire scene out of 3D objects. Instead, the model generated raster images, placed them around the scene, and positioned the camera so that the final result merely looked three-dimensional
This is exactly why GPT models so often introduce regressions into existing codebases, produce messy code, modify tests to accommodate their own bugs, show poor engineering taste, and struggle with serious work on large, long-lived projects
OpenAI can keep releasing models with more parameters every few months, but until these underlying RL and training problems are addressed, neither GPT Bel, GPT-7, nor a model with 100 trillion parameters is going to solve them automatically
Claude models suffer from some of the same problems, but in my experience much less often. Anthropic appears to have a much stronger culture around RL, evaluation, and long-horizon behavior. That's one reason their models tend to show better taste, follow user intent more faithfully, and produce code that feels more deliberate, structured, and maintainable
可燃冰🍥 retweeted
used sol 6.1 for a whole day and i'm very.. confused
will try my best to break down the key things to know
1. sol 6.1 reminds me of opus 5
this is a model that makes me constantly have to ask "wtf are you talking about"
it's very capable. most of the time once i understood what it meant, it did mean the right thing. but it's a pain to get through its choice of words
i hope there's a sol 6.5 coming that ends up being what opus 5.5 was to opus 5 and get this fixed
2. it's bigger than previous sol
you can easily observe the slowness that comes with this model. this hints at it being a bigger model than 5.6 sol. but my confusion is why it's priced at terra and sonnet range
then i realized, the pricing is set to attract enterprise customers who pay at API rate
for consumers who buy the $200 subscription, the lower price doesn't mean more usage because they're cutting the subscription's quota by half
so i guess the intention is to give enterprise customers an opus competitor at lower price, while holding back the benefit from consumers
3. all gpt models still suffer from poor judgment
sol 6.1 is likely taught by astra. if you read some of its code, it smell like astra code. it also inherited astra's strength which is that it can get stuff done with less turns and tokens (even opus 5.5 hasn't got to this level of efficiency yet)
but... the biggest problem to me is still "judgment". this is most obvious when i start to use either astra or sol 6.1 as my firstmate - everything simply starts to fall apart because of poor decision making
for example it would see me approving a plan and decide to file it in the backlog rather than the more obvious next step which is to start implementation (which opus 5.5 and fable would always "just know" it should do)
all signs point to a lack of human judgment in post training - all gpt models are likely heavily optimized by RLVR
overall - i'd categorize sol 6.1 as a solid implementer and reviewer, but i would not use it interactively
openai has work to do - right now they simply don't have an opus 5.5 equivalent anywhere in its line up, which is problematic
but seeing how anthropic could come back from opus 5, there's still hope :)






























