thorum

Interesting that all four models converge on such similar designs, for such short prompts.

Interesting read! Creating tests is highlighted as something Claude did well, but it strikes me that all the weaker rejected solutions could have been avoided if it were really good at designing intelligent tests for itself. For example, the first solution “was very specific to the reported bug and wouldn’t have fixed the general case” and the third suggestion “prevented the perfectly valid use of as conversion expressions in go commands as well”. I imagine both of these cases could have been noticed and avoided by the agent if it had planned out adequate tests ahead of time.

The “correct”, elegant way for AI to interact with existing software would take decades and billions of dollars to build. Someone would have to do the hard work of building new APIs, solving decades of accessibility issues, etc.

Or you can show an AI screenshots and ask it where to click.

The team with the most star power and hype tends to attract the best young talent.

If the next big breakthrough in AI comes from Anthropic, good chance it comes from some genius you’ve never heard of who decided to work there because of [famous researcher].

I wish them all the best and hope they succeed, but can’t help but suspect they’ve fallen into deep LLM psychosis. Even if you assume they can build this thing and it works as described and then get past all the regulatory hurdles, the scale of infrastructure they’re talking about is enormous.

I actually think “explore Claude’s understanding of colors” is an interesting concept. A lot of fascinating cultural information gets compressed into LLMs.

Unfortunately for the people mad about this, I predict the only thing they will accomplish by pressuring the rsync maintainers, is to discourage everyone else from responsibly disclosing their use of AI. You’re just going to make people disable Claude attribution on their commits to avoid drama.

Somewhat useless article. To summarize, we have anecdotes suggesting they may work but no one has figured out how to prove or disprove it in a study, and the author has some doubts. Meanwhile supplements can be dangerous if you take too much or have a liver condition, or if you buy them from an unreputable source, as with every other substance on earth. Author confirms it tastes good in milk.

I was really unimpressed by the free Codex (for nodejs/react dev). I think it must be using a less powerful model or they’re limiting it in some other way.

You’re probably right in a literal technical sense, but a very large number of people (maybe most?) would choose “no” if properly informed and asked for consent, and lots of people are morally opposed even in principle to downloading a large AI model onto their computer. I’m not one of them, but they’re out there. So in a cultural sense, it is different.

The models are primitive right now, but we’re clearly heading toward “AI as sound synthesis, human as artist” - much like how producers currently use a DAW to assemble premade loops and sounds from Splice, but with the producer now able to prompt any sound, filter, or effect they can imagine into existence and then rearrange them into a song.

See for example Suno Studio, which is not very good in my opinion, but shows the direction they’re going.

Isn’t this a permissions issue? Your “opt out” is using a GitHub access token that doesn’t allow it to happen.

I have the opposite experience: random HN/Reddit comments saying “this sucks” or “whoa this is a huge improvement” are the only benchmark that means anything. Standard benchmarks are all gamed and don’t capture the complexity of the real world.

Stars have been useless as signals for project quality for a while. They’re mostly bought, at this point. I regularly see obviously vibe-coded nonsense projects on GitHub’s Trending page with 10,000 stars. I don’t believe 10,000 people have even cloned the repo, much less gotten any personal value from it. It’s meaningless.

Good day for Kling.

Ape thinking is a cognitive practice where a human deliberately solves problems with their own mind. Practitioners of ape thinking will typically author thoughts by thinking them with their own brain, using neurons and synapses.

The term was popularized when asking a computer to do it for you became the dominant form of cognition. "Ape thinking" first appeared in online communities as derogatory slang, referring to humans who were unable to outsource all their thinking to a computer. Despite the quick spread of asking a computer to do it for you, institutional inertia, affordability, and limitations in human complacency were barriers to universal adoption of the new technology.

Their design approach wasn’t particularly unusual, so I’m not sure what that sentence means.

I do miss the days when technical reports were clear and concise. This one has some interesting information, but it’s buried under a mountain of empty AI-written bloat.

AI for help figuring things out and Timeshift for when you accidentally break something. One reboot and it’s fixed.

but the number of problems requiring deep creative solutions feels like it is diminishing rapidly.

If anything, we have more intractable problems needing deep creative solutions than ever before. People are dying as I write this. We’ve got mass displacement, poverty, polarization in politics. The education and healthcare systems are broken. Climate change marches on. Not to mention the social consequences of new technologies like AI (including the ones discussed in this post) that frankly no one knows what to do about.

The solution is indeed to work on bigger problems. If you can’t find any, look harder.

I’m honestly surprised LLMs are still screwing up citations. It does not feel like a harder task than building software or generating novel math proofs. In both those cases, of course, there is a verifier, but self-verification with “Does this text support this claim?” seems like it ought to be within the capabilities of a good reasoning model.

But as I understand the situation, even the major Deep Research systems still have this issue.

The article presents AGENTS.md as something distinct from Skills, but it is actually a simplified instance of the same concept. Their AGENTS.md approach tells the AI where to find instructions for performing a task. That’s a Skill.

I expect the benefit is from better Skill design, specifically, minimizing the number of steps and decisions between the AI’s starting state and the correct information. Fewer transitions -> fewer chances for error to compound.

Agree that planning time is the bottleneck, but

3 days

still seems slow! I’m saying what happens in 2028 when your entire project is 5-10 minutes of total agent runtime - time actually spent writing code and implementing your plan? Trying to parallelize 10m of work with a “town” of agents seems like unnecessary complexity.

Am I wrong that this entire approach to agent design patterns is based on the assumption that agents are slow? Which yeah, is very true in January 2026, but we’ve seen that inference gets faster over time. When an agent can complete most tasks in 1 minute, or 1 second, parallel agents seem like the wrong direction. It’s not clear how this would be any better than a single Claude Code session (as “orchestrator”) running subagents (which already exist) one at a time.

I agree that LLMs can be useful companions for thought when used correctly. I don’t agree that LLMs are good at “supplying clean verbal form” of vaguely expressed, half-formed ideas and that this results in clearer thinking.

Most of the time, the LLM’s framing of my idea is more generic and superficial than what I was actually getting at. It looks good, but when you look closer it often misses the point, on some level.

There is a real danger, to the extent you allow yourself to accept the LLM’s version of your idea, that you will lose the originality and uniqueness that made the idea interesting in the first place.

I think the struggle to frame a complex idea and the frustration that you feel when the right framing eludes you, is actually where most of the value is, and the LLM cheat code to skip past this pain is not really a good thing.

Your other comment sounded like you were interested in learning about how AI labs are applying RL to improve programming capability. If so, the DeepSeek R1 paper is a good introduction to the topic (maybe a bit out of date at this point, but very approachable). RL training works fine for low resource languages as long as you have tooling to verify outputs and enough compute to throw at the problem.

Go read the DeepSeek R1 paper

Developed by Jordan Hubbard of NVIDIA (and FreeBSD).

My understanding/experience is that LLM performance in a language scales with how well the language is represented in the training data.

From that assumption, we might expect LLMs to actually do better with an existing language for which more training code is available, even if that language is more complex and seems like it should be “harder” to understand.

I remember reading and hearing similar rants from programmers 15 years ago, long before LLMs. The author kept going and figured it out, and probably got some pride and enjoyment from finishing the project in spite of the frustrating moments. That’s what learning to code has always been like.

I would disagree. If you have no class in private, you have no class.

Hollywood had a ton of issues but it at least had some... class?

It looked that way because they had media training and their public personas were carefully managed, with staged interviews and media appearances. Behind the scenes, it’s a different story.

Influencers are rewarded for seeming authentic. Mr Beast coming across badly in a traditional TV interview just makes his audience think he’s more real.