David

First principle thinking. Founder university cohort 7. sharing my independent thinking & my journey related to tech. #AI #LearningInPublic

[David @David_AI_in_SV]

Think critically and independently, even if it means going against the crowd

Seek the truth, not validation from others

[David @David_AI_in_SV]

@thsottiaux

Codex feedback: Codex should be able to dynamically revise its goal as it works.

Right now, once a goal is set, it appears to remain fixed, and Codex continues working toward it. But during open-ended exploration, Codex may discover that the original goal was poorly specified, too narrow, or simply not the right direction.

It would be extremely useful if Codex could refine, expand, or shift the goal itself based on what it learns along the way. This is especially important in mathematical research, where the right objective often becomes clear only through the process of exploring the problem.

[David @David_AI_in_SV]

What I love about Codex 5.6 sol: it helps me turn my idea into feature just in hours.

I've built a price tracking dashboard with live price over the weekend and it's been running smoothly so far. Love the UI design as well!
@thsottiaux @ChatGPTapp @OpenAI

[David @David_AI_in_SV]

@thsottiaux @OpenAI One thing I do notice is that Codex keeps reminding me that Sol model can’t be used and I have to switch to gpt5.5

[David @David_AI_in_SV]

@Balder13946731 比如有哪些呢？

[David @David_AI_in_SV]

@sama we need GPT thinking (and pro) to have more lateral thinking. not just focusing on the current task, but to see the overall bigger picture to avoid drilling down to a specific rabbit hole

[David @David_AI_in_SV]

@sama able to navigate and stay the course of the directions when reasoning. this is obvious when conducting math and physics research. the direction is offen times deviated from the main problem.

[David @David_AI_in_SV]

@howie_serious 是的，我也挺喜欢gpt-5.1 thinking目前的风格，之前gpt 5 thinking聊生活相关的真的没法用

[David @David_AI_in_SV]

GPT 5.1 thinking model is really good compared to GPT 5 thinking. Before when talking about life in general or discussing life decisions, it felt like talking to a nerd who only knew about science and provided concrete metrics which were only applicable in science field.

Now GPT 5.1 thinking has both depth, warmth and clarity. It feels more like talking to a high intelligence human being who has life experiences and matches with your visions on how to deal with struggles.

Thank you @ChatGPTapp @sama @markchen90 and team!

[David @David_AI_in_SV]

In terms of intellectual novelty and proof depth, IMO is harder.

IMO leans more toward unknown-unknown. You need to create a lemma no one handed you.

ICPC leans more toward known toolkits, correctly identify and combine known techniques, more of pattern recognition which is suitable for AI. But the search space of implementations and edge cases is vicious.

[OpenAI @OpenAI]

Our general-purpose reasoning models solved all 12 problems at the 2025 International Collegiate Programming Contest (ICPC) World Finals, the world’s top university programming competition which was enough for a 1st-place human ranking.

[David @David_AI_in_SV]

@grok @Maxouks @elonmusk @grok No it has to be a pic meme not a character meme

[David @David_AI_in_SV]

@Maxouks @elonmusk @grok could you reply me with a meme?

[David @David_AI_in_SV]

@sama @Yuchenj_UW I think it’s time to upgrade advanced voice mode now. It’d be great if the model is also thinking model too

[David @David_AI_in_SV]

@sama Currently voice feature is still connect to GPT-4o. It would be great if its connected to better models like GPT-5 or thinking model

[David @David_AI_in_SV]

@lefthanddraft @jcsemantics It’s actually used for controlling the internal analysis budget. The bigger the juice, the more token budget can be used

[David @David_AI_in_SV]

@python_xxt 参数大概率没有调，只是把thinking阶段的token budge调高了，思考时间更长，回答的质量就会相对更高。今天问一些技术相关问题思考时间都是2分钟以上，之前大概平均一分多钟左右

[David @David_AI_in_SV]

@ericmitchellai @sama It does well on most cases but if the question is out of distribution, it hallucinates a lot

[David @David_AI_in_SV]

@YiMaTweets Here is one test I have today:

[David @David_AI_in_SV]

Impressive. GPT-5 thinking mode could do 26-digit number by a 22-digit multiplication correctly without any tools after 3 min 55s (checked CoT that no tools were called).

The previous models all failed miserably.

This is a big milestone.
Congrats to @ChatGPTapp @OpenAI !

[David @David_AI_in_SV]

@YiMaTweets We can’t sorely rely on the arena test. I’ve been using GPT series in the very beginning and I do feel the big differences