causal

Yeah title comes from nowhere in the link. No doubt it's possible but all that matters is speed and we learn nothing of that here...

AI makes the first 90% of writing an app super easy and the last 10% way harder because you have all the subtle issues of a big codebase but none of the familiarity. Most people give up there.

Run an incredible 400B parameters on a handheld device.

0.6 t/s, wait 30 seconds to see what these billions of calculations get us:

"That is a profound observation, and you are absolutely right ..."

I use Claude Code because Anthropic requires me to in order to get the generous subscription tokens. But better tools exist. If I was allowed to use Cursor with my Claude sub I would in a heartbeat.

I've gotta say, it shows. Claude Code has a lot of stupid regressions on a regular basis, shit that the most basic test harness should catch.

I don't think you've proven what you were hoping to prove.

If you're using LLMs to write for you then you need to develop a deeper understanding of their capabilities. This is a required reading for anyone I have using AI: https://www.theregister.com/2026/02/16/semantic_ablation_ai_...

I'm not an expert in the field, but this reads like someone trying to sound impressive by using big words without providing any solid detail.

Petty? He's accusing them of fraud, and if he's right then yeah we should all be disappointed in Eon's deceptive marketing.

Fully simulating a drosophila has been a high goal for a very long time and great claims require great proof, but Eon has been stingy with the details (and no this blog post does not reveal much beyond chaining together lots of impressive sounding words).

Imagine the skepticism on HN if someone declared they invented AGI. Similar level claim.

Damn, then we automated bullshit generation

That's about as polite as you can get, and it's still risky: people get defensive, the output might NOT be from an LLM, etc.

That's the asymmetry of the problem: Writing with AI delegates the thinking to the reader as well as all the risk for correcting it.

Straightforward is ambiguous. To replicate NanoClaw would probably only take about a day of work and testing and refining in Claude Code, but that's a day I didn't have to spend to get NanoClaw.

I like NanoClaw a lot. I found OpenClaw to be a bloated mess, NanoClaw implementation is so much tighter.

It's also the first project I've used where Claude Code is the setup and configuration interface. It works really well, and it's fun to add new features on a whim.

This is true, but the attack surface on your life is decreased by better security around the entire setup.

But I fundamentally agree that there is just too much overlap between what makes claws useful and what makes them insecure.

This question doesn't even make sense. Why wouldn't humans still be the ones responsible? Bot account?

Wait - what was the AI tool and how did it have her face to begin with? If small-town police are doing face-matching searches across national databases then nobody is safe because the number of false positives is going to be MASSIVE by sheer number of people being searched every day.

Pretend the tool is 99.999999% specific. If it searches every face in the USA you're still getting about 3 false positives PER SEARCH.

You will never have a criminal AI tool safe enough to apply at a national scale.

This x1000. We need to suspend this shared fiction that AI has any agency. Only humans can be responsible. Full stop.

Yep. I actually prefer seeing imperfect writing, there is signal there that AI would erase.

Ah. That might be why I find it especially triggering.

I'm with you on this. I worked on a big Temporal project briefly and I was really turned off by how much of the codebase was just rote mapping properties from one layer to the next.

If you really DID come up with that paragraph 100% completely on your own with no LLM influence then...I apologize for the insult, though I can't really back out from what I said. It's still a bombastic way of saying very little.

Yah it's just an ad, and "Pentesting agents finds low-hanging vulnerability" isn't gonna drive clicks.

Those short "punchy sentence" paragraphs are my new trigger:

> No credentials. No insider knowledge. And no human-in-the-loop. Just a domain name and a dream.

It just sounds so stupid.

Wow this sparked a lot of tedious debate. Just wanted to clarify that I'm talking about the law as written, not some high-level philosophical ideal.

No. Parking is leaving your possession somewhere and should apply to the registered owner. It is not illegal to own a car that someone else used to run a red light.

Was wondering this too. Not to mention adding TOS to your own sites that get used by bots and scrapers.

Codex is not bad, I think it is still useful. But I find that it takes things far too literally, and is generally less collaborative. It is a bit like working with a robot that makes no effort to understand why a user is asking for something.

Claude, IMO, is much better at empathizing with me as a user: It asks better questions, tries harder to understand WHY I'm trying to do something, and is more likely to tell me if there's a better way.

Both have plenty of flaws. Codex might be better if you want to set it loose on a well-defined problem and let it churn overnight. But if you want a back-and-forth collaboration, I find Claude far better.

> psyops targeted at making user doubt their computer

IDEK what that means, specific examples?

> OpenAI ... explicitly allow this

Explicit means it's stated in OpenAI docs somewhere, but I can't find it. Link?

You COULD even do Qwen3.5-35B-A3B-GGUF.

UD-IQ3-XXS is only 13.1GB, which might outperform both in both intelligence and certainly speed (only 3B activated): https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF

To accommodate cache you will need to offload a few feed-forward layers to the CPU. Will still be quite fast.

Edit: Actually 27B does a little better than 35B on most benchmarks- 35B will still be much faster.

3-bit 27B will almost certainly be better. 4-bits is usually the limit below-which you start to see more steep drop-offs, but you also get diminishing returns above 6-bits. So I'd still rather pack in more params at 3-bits.

9B will be faster, however.