There are only two important engineering problems in the world. 1) Probability density estimation in high dimensions.2) Solving high dimensional optimization problems. Both problems are not solvable in the general case (curse of dimensionality, NP hard). But in practice,.

Unpopular opinion: if we really had a real jurassic park, as in a zoo for dinosaurs, people would not get tired of visiting it after 30 years and the need to genetically engineer new species would not exist . Although engineering new dinosaur species sounds great, actually.

I still kinda bogles my mind that the CEO of a ~$40B robotics company is acting like a twitter-influencer-threadboy in the last ~2 years with these posts. very weird behavior.

looking forward to the deep research and agent updates based on this model. there might be a usefulness leap there.

I like the fact that GPT5-thinking indeed appears to be hallucinating much less compared to o3.seems more trust worthy. it's also is extremely good at searching the web. These two properties combined make it the best search product out there by miles - saves me a ton of time.

The "thinking vs. non thinking" router is a weird psyop.nobody should ever prefer to have thinking off even for "easy questions". just keep it always on and and let the reasoning model itself decide when it has been thinking enough for a sufficient answer.

It is becoming clear that openai is responding to market demans and will try to make model nicer and more polite (and more sycophantic), etc. This is OK as long as this is not in any way harming the focus on making the models smarter or more capable. its a classic case of users.

I'm sure that o3-like RL post training on GPT4.5 base would have yielded much much better results, but would have been cost prohibitive and thus not interesting from business perspective.

or perhaps the correct way to think about it is that the business culture, that prioritizes usability, has finally won over the research culture, that prioritizes achieving new heights, in openai?.

Practically, I hope that this base is strong enough to create better versions of deep research, codex and agent, and that they could think for much longer at similar costs and thus achieve better outcomes.

so the tl;dr for GPT5 is that it doesn't move the frontier of intelligence, but gives us the frontier at reduced cost. still a positive development, but everyone expected that the frontier would move as well at least a bit, and for all paid users there is no change in day to day.

instead of using it's "eyes" to see, it tried to vibe code a crappy detection algorithm by detecting "orange" colored blobs in the image. over-eager to use tools. further instructions didn't yield any better results

a clearly better model, even if by a bit, will get all the usage (increases market share). and for unpaid subscribers this could be a big driver for much increased usage. in this sense, a cheaper model increases the market size. I suspect anthropic will see very substantial and.

based on the live stream, it appears to be even more incremental than expected.

livestream became boring, so came here.first 20 posts in feed are about the the fact that 52.8 > 69.1 = 30.8.

its clearly the case that some people are just much better at using AI models to help them do things than others, but its not clear what is the reason for this. feels like something about resourcefulness and ability to think before doing (as opposed to thinking while doing), but.

don't get me wrong, small increments eventually add up, but this just screams "we trained it 10% longer".

This should be the oxford dictionary definition of "incremental"

entire world simulation inside the weights and activations of a neural network. amazing.

this is the way!. deepmind for the win?.

40 years ago, if you wanted to go somewhere you had to ask someone for instructions, if you wanted to find a lawyer or a doctor, you needed to ask multiple someones if they know anyone, if you wanted help of any kind, you had to ask someone. basically, information was either.