WiSaGaN

Maybe I should start a project rewriting pctools 5.0 in rust!

Yeah, this seems like a personal choice, which does work out given the current result.

Harness is fine. I think people here are arguing what provided here to take the test is not harness.

This violates the ToS, but I don't think it's distillation. Distillation requires knowing the logits, which current API does not provide. This is just synthetic data generation. Anthropic definitely knows the difference.

Google has gigantic power over its users. Consider that for some reason, Google banned your gmail account, which you are using for large number of logins for different essential services.

I think diffusion makes much more sense than auto-regressive (AR) specifically in code generation comparing to chatbot.

Great article. To me, this highlights a key question in the era of rapidly advancing machine intelligence: if we know machine intelligence is progressing, what is more valuable to build for? As humans, we still find many tools useful even when doing knowledge work. For instance, a calculator. Sure, a smart person can perform calculations in their head, but it’s much easier to teach everyone how to use a calculator, which is 100% reliable in its intended domain.

In this era, we should build these kinds of tools for problems we know are straightforward ones you can’t get smarter than, even as intelligence continues to advance. Using tools like "bash" or command-line interfaces originally designed for humans is a good initial approach, since we can essentially reuse much of what was built for human use. Later, we can optimize specifically for machines, either accounting for their different cognitive structures (e.g., the ability to memorize extremely long contexts compared to humans) or adapting to the stream-based input/output patterns of current autoregressive token generators.

Eventually, I believe machine intelligence will build their own tools based on these foundations, likely a similar kind of milestone to when humans first began using tools.

Yes, agentic-wise, Claude Opus is best. Complex coding is GPT-5.x. But for smartness, I always felt Gemini 3 Pro is best.

OpenAI has a former NSA director on its board. [1] This connection makes the dilution of the term "PRISM" in search results a potential benefit to NSA interests.

[1] https://openai.com/index/openai-appoints-retired-us-army-gen...

Rust's `serde_json` recently switched to use a new library for floating string conversion: https://github.com/dtolnay/zmij.

A market maker needs a premium to provide liquidity. If all else is equal, why would they take on execution time risk? This is a universal feature of continuous-trading Central Limit Order Books (CLOBs), not something unique to prediction markets.

I guess I am asking how we know Gemini and Claude relies on the additive residual stream. We don't know the architecture details for these closed models?

How do you know "GPT-5, Claude, Llama, Gemini. Under the hood, they all do the same thing: x+F(x)."?

I have a similar experience when I found out that claude code can use ssh to conect to remote server and diagnose any sysadmin issue there. It just feels really empowered.

I think it's likely someone already discovered this. It's just that info is not broadcasted to people who want to comment on this thread.

I can't find the source. Anyone can point to it?

This argument falls apart when you look at Rust and Cargo. uv is literally trying to be "Python's Cargo." The entire blueprint came from a flagship FOSS project.

Rust's development used a structured, community RFC process—endless planning by your definition. The result was a famously well-designed toolchain that the entire community praises. FOSS didn't hold it back; it made it good.

So no, commercial backing isn't the only way to ship something good. FOSS is more than capable to ship great software when done right.

It is now. But the limit on $20 plan is quite low and easy to use up.

Investment banks in Hong Kong were almost exclusively western back in the days with very few ethnic Chinese in senior management.

China is not the birthplace of so called '996'. Long before tech scene in China, there are a lot of investment banks doing that in HK especially for junior analysts. Calling 996 a China thing is just orientlalism. Everything bad is Chinese, everything good is western.

That's true. You would think LLM will condition its surprise completion to be more probable if it's in a joke context. I guess this only gets good when model really is good. It's similar that GPT 4.5 has better humor.

This actually makes sense because in the meantime Meta is ditching the open-source (open-weights) direction.

Before the national security narrative took over, the main argument was about "safe" AI, where releasing models as open weights was considered "not safe." Now that no major US AI players release premium open-weights models, the "safety" narrative isn't needed anymore—so cooperating with the US military is feasible again.

I am wondering how much of this can be mitigated by carefully designing feature flags, and make default feature set small.

Can you name one example of a consumer product that China initially sold affordably to gain market share and then later raised prices?

Deepseek and Alibaba just published their froniter models in open weights weeks ago. And they happen to be the leading open weights models in the world. What are you talking about?

Deepseek published thinking trace before OpenAI did, not after.

I don't think mentioning Rust on an article specifically talking about a memory safety bug count as "constant". This is Rust's core strength.

Interesting that the output price per 1M tokens is $0.6 for non-reasoning, but $3.5 for reasoning. This seems to defy common assumption of how reasoning models work, and you tweak the <think> token probability to control how much thinking it does, but underlying it's the same model and the same inference code path.

This common feature requires the user of the API to implement the tool, in this case, the user is responsible to run the code the API outputs. The post you replied suggests that Gemini will run the code for the user behind the API call.

Also the "project" feature in claude improves experience significantly for coder, where you can customize your workflow. Would be great if gemini has this feature.