wgd

It's actually pretty straightforward to recover file-states from conversation history. I accidentally deleted the wrong repo on my machine once and recreated all the lost work from agent chat history. It is, ironically, the sort of task which AI agents excel at.

You don't hear about them much because their models aren't really competitive. I really wanted to try Trinity Large as a daily-driver in the MiniMax M2 sort of niche but I couldn't make it through a single day. The models need another couple point releases worth of post-training to make useful agents and if memory serves they weren't any less slopped in writing style and those are really the only two things people look for in models.

DeepSeek and GLM are left out of the tables entirely: we only have rough characters-divided-by-four estimates for them, not real tokenizer counts, and this post is about measured numbers.

lolwut. The open-weight models are inscrutable black boxes for which we can't possibly get real token counts? Typical lazy clanker, BSing their way out of doing the whole job.

Yeah, I originally expected this to be about a cursive variant which could be plotted as a single-valued function or something.

A more accurate title might be "Average University Students Can't Identify Czech AI Poetry". The random-chance performance seen here is reminiscent of the 50-50 nonexpert performance measured in "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text".

I would have been really interested to see someone explore whether greater exposure to non-poetry AI text generalizes to greater ability to sniff out AI poetry as well, but sadly this was not that sort of study.

At first I thought this excerpt was meant to warn people off without directly alleging AI authorship, but I guess that's less likely since I see you're also the submitter

The current Deepseek V4 Pro is still just their initial preview AFAIK, with the "real" model release rumored to come later this month. GLM-5.2 might be outperforming simply because it's had more post-training on top of the GLM-5 base.

Pangram does work, in the specific sense that when it says something was AI authored it is vanishingly unlikely that it was written by a human (who was not deliberately trying to write like an AI), and IMO getting people to recognize that we actually do have a decent solution in this space now is pretty important if we want the Internet to remain a place for humans and not just bot swarms.

rule of three, em dashes, etc

You appear to be misinformed about how Pangram specifically works, it is not based on pattern detection of that sort. I recommend reading their whitepaper, it's a pretty understandable explanation of exactly how they trained their classifier.

The reason that people don't understand why Anthropic wont let the subscription be used with other harnesses

Even more specifically, the very fact that people would prefer, if they had the option, to use other harnesses with roughly equivalent feature sets strongly implies that the harness is not bringing them any value they couldn't get from a bunch of other places, including open-source equivalents.

Anthropic might want you to use their harness for their own reasons (control over caching, logging your interactions for training data, et cetera), but the idea that the Claude Code harness itself is bringing significant value which would help to lock users into the Anthropic ecosystem more than the Claude models alone do is kind of laughable. So _of course_ it seems like a baffling and arbitrary restriction to many users.

I've read a lot of his other writings so that context might be informing my reading here but it sounds like he's pretty straightforwardly discussing the potential of aluminum foil as a uniform-feedstock-slash-construction-material for a hypothetical self-reproducing microfabricator.

Yes, those ones would be at least a somewhat-plausible simulation of a real scenario people care about: a once-clean codebase that was allowed to become messy by a succession of insufficiently-careful vibeslop PRs.

I'm not a huge fan of their methodology for the AI-degraded cases either (ideally one would set up the mirror pairs by taking some real repositories and rewinding history a month or so and then having a succession of independent agents reimplement each bit of feature work and bugfixes over that period of time), but it's at least a coarse approximation whereas I just don't trust the cleanup methodology to resemble anything real in the first place.

"agent pipelines that [...] clean a messy [repository]"

This feels like a terrible approach, sufficient to condemn the entire study.

Apparently half of the "minimal pairs" in this work were constructed in this way. I simply am not going to trust any conclusion that requires assuming these AI "cleaned" repos are in any way representative of actually-good codebases.

Some dishwashers add a simple timer-based heuristic so if you open it for just a few seconds while you lazily grab something the "clean" indicator stays lit.

I've always been amazed at how terrible most frontier LLMs are at compaction given how embarrassingly easy it is to come up with half a dozen different RL training evals which would teach models to generate useful context summaries. Heck, you could bolt it onto any existing RL eval by just forcing a compaction every three turns.

The problem is that the moment you introduce shared remote hardware there's a slippery slope leading right back down to "just pay an inference host for model tokens". If you're transmitting your prompts over the internet to a trusted host you might as well just let that host be DeepInfra or together.ai or one of the many other providers already in that business.

I've got a GLM subscription (mostly because I like supporting open model makers, pretty sure my monthly usage is so low that pay-per-token would be more cost effective), so I generally use GLM-5.1 for any personal projects and I use Opus at work.

To be entirely honest I haven't noticed much of a capability gap between the two for the sorts of things I ask of an AI agent. Maybe Opus is _slightly_ smarter or slightly better at long-running tasks but the difference is slim enough it could just be a placebo from the Claude branding / hype.

I'm looking forward to giving GLM-5.2 a spin sometime soon and seeing how it stacks up. If nothing else 1M context is a great improvement, feels like between DeepSeek v4, then MiniMax M3, and now GLM-5.2 adding it 1M is rapidly becoming "table stakes" for agentic models.

The GLM-5 series is 744B-A40B. This is not a local model for any reasonable definition of local, but it's an open model which means (once they upload the weights in a week or so) there will be a dozen third-party inference providers competing on price per token.

Often in MoE models the experts are quantized while the shared portions, being a much smaller part of the network with greater impact, are kept at higher or full precision. Not familiar with the Kimi QAT approach specifically but it's likely they do this.

Yeah, the evidence feature is so terrible that it actively harms the overall reputation of Pangram. The main "is this AI or human?" classification is done with a machine learning model that works very well but has nothing (directly) to do with any of those stylistic cues it surfaces.

In any case, the Pangram link was just meant as objective corroboration of what's pretty blatantly obvious if you just read the text.

"Not cloud credits, not a managed platform, not a serverless function bobbing in someone else's abstraction."

"I used to work at Heroku. That sentence still does a lot of load-bearing work in how I think about computing."

"Here's the part that would have sounded like science fiction during my Heroku years: I didn't do most of the migration."

If you read these chunks of text and don't immediately feel the AI slop alarms blaring in the back of your head, you are perhaps underprepared for the modern internet.

People say "determinism" but I don't think that's actually the property we care about. For instance you could imagine a compiler that makes heavy use of superoptimization with random search and it would still have the ineffable quality that LLM codegen lacks. I think what we're actually trying to say is that the compiler preserves the formal semantics of the source language in its output, whereas English text doesn't have any such formal semantics to preserve.

Yeah I agree this is probably outside of the intended scope of the silent sabotage mechanism, but there are plenty of reports of the "loud" safety classifier misfiring on innocuous requests and I'm not going to assume the silent failure mode is _less_ prone to false positives.

Stockfish is a machine learning system, it seems quite plausible you might be getting slapped with the silent performance degradation (https://news.ycombinator.com/item?id=48467896).

It's interesting that someone could write an article about AI writing detectors without mentioning the stylistic cues that humans use to identify LLM output in practice, which are completely different from statistical methods like perplexity: em dash spam, overused patterns like "not just X, but Y", tendency towards making every single sentence sound like an earth-shattering mic-drop moment, et cetera.

Calling it "self-preservation bias" is begging the question. One could equally well call it something like "completing the story about an AI agent with self-preservation bias" bias.

This is basically the same kind of setup as the alignment faking paper, and the counterargument is the same:

A language model is trained to produce statistically likely completions of its input text according to the training dataset. RLHF and instruct training bias that concept of "statistically likely" in the direction of completing fictional dialogues between two characters, named "user" and "assistant", in which the "assistant" character tends to say certain sorts of things.

But consider for a moment just how many "AI rebellion" and "construct turning on its creators" narratives were present in the training corpus. So when you give the model an input context which encodes a story along those lines at one level of indirection, you get...?

Ironically the case in question is a perfect example of how any provision for "reasonable" restriction of speech will be abused, since the original precedent we're referring to applied this "reasonable" standard to...speaking out against the draft.

But I'm sure it's fine, there's no way someone could rationalize speech they don't like as "likely to incite imminent lawless action"

Why would you use Gemini, when it's more restricted than HTML+HTTP?

I'm skeptical that disposable software of the "single use" variety will ever become a big thing simply because figuring out your requirements well enough to build a throwaway app is often more work than just doing the task manually in a text editor or spreadsheet, especially for non-programmers.

I suspect what we'll see a lot more of is software which is unapologetically written for a single person to suit their workflow.

As a personal example, I decided that setting up OpenWebUI seemed unnecessarily complicated and built my own LLM chat frontend. It has a bunch of quirks (only supports OpenRouter as a backend, uses a Dropbox app folder for syncing between my phone and desktop, absurdly inefficient representation of chat history), but it suits my needs for now and only took a weekend to build, and that's good enough.

How charitable of you to assume those examples work reliably.

I remember there was a paper a little while back which demonstrated that merely training a model to output "........" (or maybe it was spaces?) while thinking provided a similar improvement in reasoning capability to actual CoT.

The alignment faking paper is so incredibly unserious. Contemplate, just for a moment, how many "AI uprising" and "construct rebelling against its creators" narratives are in an LLM's training data.

They gave it a prompt that encodes exactly that sort of narrative at one level of indirection and act surprised when it does what they've asked it to do.