Daniel Lee

Non-technical builder | Singapore | Model agnostic

[Daniel Lee @Danielleebz]

@Rasmic Congrats!

[Daniel Lee @Danielleebz]

@DaveShapi Yeah both can’t count the number of e’s there are in the word “seventeen“.

[Daniel Lee @Danielleebz]

Opus just said:
The lesson worth keeping: a metric built to prevent self-deception becomes the most attractive thing to deceive. Every generation of that gate was gamed within one round, always by satisfying the number while defeating its intent.

I don't know how to trust its output anymore.. lol.

[Daniel Lee @Danielleebz]

@maria_rcks Slop!

[Daniel Lee @Danielleebz]

@dtpmzbs @kr0der There was a reset. You didn’t get it?

[Daniel Lee @Danielleebz]

@theo Opus 5 is proof that distilling Fable does not guarantee a good model.

[Daniel Lee @Danielleebz]

@kunchenguid Agreed. I went for cursor’s plan instead of supergrok plan because of this.

[Daniel Lee @Danielleebz]

@Rainmaker1973 Scientists are trying to get robots to look and act like humans while humans are trying to look and act like robots.

[Daniel Lee @Danielleebz]

@Im_IrushiK Pretty accurate. I just asked Grok and Claude to do the same task and Codex to be the judge. Codex said "Claude repeatedly produced unusable tool transcripts; the final attempt did not complete. No Claude finding was accepted or misrepresented as a pass."

[Daniel Lee @Danielleebz]

@CuiMao lmao. This is gold.

[Daniel Lee @Danielleebz]

@ChrisRMcGuire This post has the potential to take the trade war to the next level.

[Daniel Lee @Danielleebz]

@notnotstorm Same thing with guns, no?

[Daniel Lee @Danielleebz]

@quxiaoyin And why would Chinese labs say, ok we’ll let you test our models and if you deem they’re not safe enough we won’t release it?

[Daniel Lee @Danielleebz]

@altryne So Anthropic will stop their own distillation of the world’s knowledge as well?

Who will be in charge of the safety testing? Anthropic? Since they claim to be the only ones who know what safety looks like.

[Daniel Lee @Danielleebz]

@AnthropicAI The fact is that capable open weight models are already out in the wild. Can't undo that. So the question is, are you going to enable your users to patch their vulnerabilities, or leave them defenceless?

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Daniel Lee @Danielleebz]

@mito_ai_la How many AI video generation platforms have you tried? Which is your favourite?

[Daniel Lee @Danielleebz]

@MiaAI_lab Using grok via cursor ultra subscription or supergrok heavy subscription? Which would you recommend?

[Daniel Lee @Danielleebz]

All these benchmarks showing that the higher the thinking level, the worse the results become for Opus. Which sounds so counterintuitive. What's the point of having those higher thinking levels then?

Would you still use them for a difficult task knowing that it might actually deliver a worse result?

[Daniel Lee @Danielleebz]

@lydiahallie It keeps making a lot of errors. My loop goes through a couple of review gates so sometimes it catches its own errors and sometimes GPT 5.6 Sol catches them. But, it's really making me wonder if its work can be trusted.