docheinestages

My gut feeling tells me it has something to do with Cloudflare. Along with AWS, they're two of the main suspects in such incidents.

All I see is slop.

Is this a compressed and retrained version of GLM-5.2?

I mean, I'm interested to know if the frontier models also get to see all questions at once. Then it's more fair game than if they just see one question at a time.

It's what happens when you don't do market research.

Isn't this cheating? Or rather, are frontier agents only looking at one question at a time? If I understand correctly, you're looking at all the examples of the exam questions. If the exam was adjusted so that you can only look at one question at a time, you won't get 44% anymore.

I don't think writing is an "AI-complete" problem. It's just that the models are inefficient at it right now, due to training and/or architecture. Good writing requires thinking, reflection, and doing multiple passes. This currently translates to using a high reasoning effort and burning lots of tokens, but big AI labs like Anthropic are struggling with handling the load, so they're doing the exact opposite and finding whatever cheap trick they can to reduce token usage. One such trick is condensing ideas into as few tokens as possible, which in my opinion is one of the reasons Claude sucks so much at writing.

Memory is not necessarily a good thing. What we need is a combination of sufficiently large context windows (10-100 million tokens) along with curate, bloat-free data.

> How can I judge what is a good memory to store? How can I avoid filling my memory with crap?

> This is a common fear with memory systems but doesn't really apply to memoryfields. Irrelevant material is simply never surfaced by the semantic search.

This is so wrong. The Achilles' heel of this approach is the RAG. What makes it worse is having lots of memories that are outdated, wrong, hallucinated, or irrelevant.

Nothing beats curated data. Memory should be regularly reviewed, compacted, and cleaned up if it's no longer valid.

Looks like the repo teleported.

This is only a problem because people keep using X. Just stop.

But did it produce anything meaningful? If so, show me.

Oh fantastic! It was already subpar and they want to make it even worse. One day we'll look back at history and see how Anthropic went down.

Kudos to Kagi!

Would it work with "giggity" too?

Can we get a tool that does the opposite for the readers?

I think one of the best use cases of AI is as a natural language interface to programming. The syntax of a programming language is an opinionated part of its design that often adds to the complexity of learning the language.

I watched the intro video and all I saw were numbers in an Excel sheet. That doesn't make this fun. I'd rather have AI ELI5 it to me.

Anthropic should build a harness (and model) that smartly takes care of all these points. Not requiring the user to do the manual work. All I see are excuses because they cannot handle the load and enforce strict quotas on users, all while OpenAI constantly resets their quotas.

With Qwen 3.8 27B, we're one step closer to on-device LLMs that can replace subscriptions.

Are the benchmarks comparing just the models while keeping the harness the same (the open-source Toast harness)?

My uses cases are mainly research and prototyping, often starting from scratch in greenfield projects. The 5-hour quota is shared between Claude web and Claude Code, so it doesn't really matter what interface I use.

When it comes to research, my prompts are already narrowed down to specific topics, and I even include examples and break the process down into stages. For development tasks, I try to avoid a mono-repo in the beginning and develop modules before combining them together to avoid distracting the AI's attention and minimize the overhead.

With Codex, on GPT-5.6 Sol with xhigh effort, I need to go several rounds and at least 2-3 hours before hitting the (now-removed) 5-hour limit, which translates to 10% of the weekly usage. In contrast, I run out of quota even with Claude Sonnet.

In terms of quality of output, Codex digs deep for research tasks, in the right direction, produces less AI slop, and follows my direction better. At least that's how I perceive it. But again, the main problem with Claude is running out of quota in the middle of research or implementing a task.

Claude has essentially become useless for agentic development or research. Doesn't matter what model you use. A few rounds and bam, you've burned through your quota. Doesn't matter how "intelligent" their models are, if you can't use them. That, and the quality of AI responses are, in my opinion, significantly worse than competitors like OpenAI. At this pace, I foresee Anthropic becoming the next Nokia.

If you would've asked me this a year ago, I would've said the exact opposite.

YCombinator's investments are unfortunately very questionable nowadays.

Your landing page is very hard to read. The font size is literally 10px for some content, while animations distract the reader.

Dear YC, please force the startups to dedicate some of the 500k funding for standard web design.

They conveniently decided not to include the Qwen range of models in the Artificial Analysis graph, except the out-of-league Max variant. At least be brave and honest.

I'm interested to learn from the prompts you used to generate the images and texts. Perhaps a skill you could share to generate such content?

Sounds like a side-effect of being VC-backed. Like what the ex GitHub CEO went on to build (Entire). Unnecessary.

Suggestion: showing an actual screenshot or video of your app is a much better indicator of effort than a generic Claude made animation. I've seen AI slop landing pages on far too many YC-backed startups. Not saying yours is one, but parts of it smell.

If you can afford it, why not. For certain phases of projects like a proof-of-concept, you need to move fast and validate several ideas. Once it's locked down, rewrite from scratch, and here, if you can afford it, type or write the code manually.

Exactly this.