Smaug123

You can literally taste the difference between wearing a mask on the Tube and not.

The article contains Dynomight’s thoughts on the literature of stylometry. Search on the word “stylometry”.

That bet isn't one anyone will take the other side of, surely - how is your counterparty supposed to collect if they win? I guess you'd propose effectively supplying a loan with a massive interest rate?

Note that Neel Nanda replicated the results on a Qwen model.

The model itself, sure; the comment is about the production of more advanced models (to keep open weights near the frontier).

By the way, you've seen Cerebras? It's not gone as far as what you described - loads of cores and RAM but you still load up the weights onto it as software and they need to be streamed into the chip for large models - but it is a whole wafer.

For what it’s worth, Claude did this without even being asked when I had it implement /dev/urandom in my deterministic dotnet runtime. (Fun fact: if the runtime only ever receives zero bytes from /dev/urandom then it will hang on attempting to initialise System.Random! That was the first way I asked for it to be implemented.)

Does it have a terse syntax? I main F#, and when I have to work with Python I generally find myself complaining about how verbose it is. (Needing intermediate variables for what should have been a pipeline, the ceremony around parallelism, having to store constructor parameters as object fields, etc.)

I didn't want to reimplement all the assembly-reading nonsense that comes for free with System.Reflection.Metadata. The `dotnetdll` crate exists but is GPL. Also in F# I can fall back to the CLR for fiddly things I don't want to implement (like the arithmetic opcodes on floats or whatever).

To answer your question, although I would certainly have preferred you to phrase your comments less insultingly: this project would otherwise never have got to a state where it could find bugs. I am not paid to write this code, and it would have taken far more years than I would have been willing to spend.

It's not actually unheard of for people to pay other entities to build their passion projects. For example, I visited [Eltham Palace](https://en.wikipedia.org/wiki/Eltham_Palace) last weekend, which was not in fact built entirely by the two Courthaulds who commissioned it.

I mean, there is a reason the MIT licence contains these words:

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND… INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF… FITNESS FOR A PARTICULAR PURPOSE…

If you would like a tool built with my organic artisanal human fingers, then I am certainly open to sufficiently large offers of money to build one for you! Alternatively, you can simply not use it if you think it won't fit your needs :)

I think the Stroop effect ("read these colour names, each written in a different colour") is probably the purest demonstration of this. Humans are trivially prompt-injectable.

A few years ago, as you say, this was true. Nowadays I guess you just have to bite the bullet that Erdős problems aren’t interesting.

Something of a passion project. It's going to fail horribly if you try and use it, I'm sure, but it can already do some neat stuff!

I take a very dim view of slopping out 500kloc and then giving it to unpaid experts to perform the actual work of checking it (confirmed at https://leanprover.zulipchat.com/#narrow/channel/583336-Auto... that this is what they did), especially given the reported incorrectness of the code (https://leanprover.zulipchat.com/#narrow/channel/583336-Auto... or https://leanprover.zulipchat.com/#narrow/channel/583336-Auto... for example).

They say in the Lean Zulip thread that this is actually intentionally a "low quality" release (https://leanprover.zulipchat.com/#narrow/channel/583336-Auto...); the paper notes that the quality is "inferior to that of expert-written Lean code". Then again, "Our results suggest that formalizing the core textbook infrastructure of modern mathematics is within reach".

("If grown, then unpredictable" is unrelated to your apparent attempted refutation "But X is unpredictable and not grown; checkmate".)

I'm not, but is it yet possible to use the Vision Pro as a window manager for macOS? I'd totally get one if it were possible to lay out ordinary macOS windows in space rather than being confined to the simulation of a single rectangular screen.

It’s still not clear to me that humanity was ready for GPT-2! Quite a lot of people claim to hate and fear LLMs. https://www.kcl.ac.uk/news/one-in-five-britons-think-ai-will... or https://yougov.com/en-us/articles/54762-most-americans-say-a... for example.

By the way, you might be interested in looking up “blameless post-mortems” and indeed the field of incident response more generally. Modern incident response practice is to treat failures of an individual to do something as problems with the system they were operating in, because humans aren’t designed to be consistent or perfect and therefore shouldn’t be pretended or assumed to be.

I think it's more that the requested information is prominently featured in the article, and indeed is the content of the only graphic in the article below the intro banner.

So far, Mythos Preview has found what it estimates are 6,202 high- or critical-severity vulnerabilities in these projects (out of 23,019 in total, including those it estimates as medium- or low-severity).

1,752 of those high- or critical-rated vulnerabilities have now been carefully assessed by one of six independent security research firms, or in a small number of cases by ourselves. Of these, 90.6% (1,587) have proved to be valid true positives, and 62.4% (1,094) were confirmed as either high- or critical-severity. That means that even if Mythos Preview finds no further vulnerabilities, at our current post-triage true-positive rates, it’s on track to have surfaced nearly 3,900 high- or critical-severity vulnerabilities in open-source code

I don't think you've addressed the fact that they can do long tasks that aren't in the training set? (And the fact that they're just statistical models isn't very relevant. So am I!)

Nah - for example, AIXI is so inefficient that it's literally uncomputable, but it is beautiful.

I think you may have read a different article from me. The thesis of the article is summarised at the end:

But if someone claims that the trend toward [X] will never reach some particular scary level, then the burden is on them to explain either:

If they’re not treating [X] as a black box, and claim to be modeling the dynamics explicitly, then what is their model? Have they calculated the obvious things…

If they are treating [X] as a black box, why isn’t their default expectation based on Lindy’s Law?

Like, the whole point is that in real life we do actually know things about situations and can model them; we fall back to Lindy's law when we know nothing at all. Further, arguments have justification to deviate from Lindy only when they give specifics about the situation they're modelling.

(If you like, you can ask an LLM what he thinks. They're all deeply familiar with his work and can certainly summarise it for you.)

Have you used the models, out of interest? They routinely do things autonomously that are not in the training set that would take me 8h, and I wouldn't say I'm slow. The profile of tasks they can do this way is jagged, and maintaining architectural coherence ("months, not hours") is still beyond them, but they're perfectly capable of writing plans and sticking to them.

Main project is a deterministic .NET runtime (https://github.com/Smaug123/WoofWare.PawPrint). Today I upgraded it to net10, which has naturally caused dozens of regressions which Claude is beavering away at.

Side project is my own agent harness, https://github.com/Smaug123/writ , which is being built sandbox-first and with Nix as a first-class citizen. Obviously everyone has to write their own agent harness as a rite of passage.

Tim Gowers is a Fields medallist who has supervised 13 students (https://www.genealogy.math.ndsu.nodak.edu/id.php?id=67729). He's totally capable of gauging what a piece of PhD-level research is.

I asked three friends to reproduce it on the same day I did, and they all reproduced it. It does seem to vary over time, which is quite spooky - although the date is part of the Claude system prompt, so one could expect some variance, I guess.

(What you call an "objective fact" here is - as you say - your report of your personal experience. Everyone else would probably use a word more like "subjective".)