@pbloem@sigmoid.social

· 7mo ago

Explaining the fediverse is simple. It's social media for people who hate social media.

There aren't that many people on it? Feature.
It's not that addictive or even engaging? Feature.
You can't make money on it and it won't help your career? Feature.

260

13

152

1

Peter Bloem @pbloem@sigmoid.social

· 3mo ago

Replying to

@RikerGoogling@mas.to Let go of the pringles, Will.

9

1

0

Peter Bloem @pbloem@sigmoid.social

· 6mo ago

Replying to

@fesshole And under hobbies they listed unicycling. That's a very inefficient mode of transport! Totally unqualified.

15

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

@ZachWeinersmith If you want to read one cover to cover I can recommend https://linear.axler.net/ It also depends a bit on what you want to get out of it. Axler is good to get the more abstract perspective on linear algebra. If you just want a concrete understanding of what an SVD does to a grid of numbers, it's probably not the best one.

linear.axler.net

8

6

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@Daojoan We've added iceberg interaction to our experience portfolio.

6

0

Peter Bloem @pbloem@sigmoid.social

· 4mo ago

I think this will be my last MacBook.

3

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Found on Bluesky (by @hannahshelley.site)

2

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

@ZachWeinersmith @lavi I think standard textbooks like Strang focus too much on row reduction as a building block and too little on the idea of a matrix as a transformation of space (which is the most intuitive starting point for me). I don't know of any textbook that does that but (if you'll forgive the plug) I wrote a long-from exploration of PCA that takes that perspective and touches on most aspects of concrete linear algebra https://peterbloem.nl/publications/unraveling-pca

peterbloem.nl

2

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

We'll see a big glut of vulnerabilities found in open source at first. That'll spook people, but it's just the inoculation period. Afterwards, open source software will be safe again, and every piece of closed source software will be a ticking time bomb.

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

So, copy.fail was found with one hour of AI assistance, and would (according to this article) have earned $500K on the open market not too long ago.

https://www.bugcrowd.com/blog/what-we-know-about-copy-fail-cve-2026-31431/

I'm no security researcher, but this kind of contradicts all those people who said that the OpenBSD bug that Mythos found (for $20K of compute) was just fancy fuzzing, and the only reason it was there was that nobody was investing 20K in OpenBSD security and the security threat of modern AI was all hype.

Bugcrowd

What we know about Copy Fail (CVE-2026-31431) | @Bugcrowd

Learn how Copy Fail (CVE-2026-31431) enables Linux privilege escalation and container escape, who is exposed, and the mitigations defenders should take.

1

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

@drgroftehauge She says on X she started a private chat. I think it's just that the description is about as vague as a horoscope. You're right that It's not really cold reading. I think ChatGPT is doing its best to apply the "rules" of palm reading. It's just that the whole language of things like this is set up to be that combination of vague and positive that make people eager to accept it whatever comes out.

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

We're really speedrunning this episode of The Scary Door, aren't we?

https://www.youtube.com/watch?v=RSEo64cW3mc

1

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

In short, driving in the real world is AI complete. You need a full, functioning brain, and then you need that brain to be trained to respond reliably. We are approaching full functioning artificial brains (yes we are), but we have no idea how to make them reliable, and they're not how current self-driving works (thank god).

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

The real world is a place of strange exceptions, especially at scale. A confused senior will get onto the highway on their mobility scooter. Children will decide to play hide and seek in a parking garage or to bike to grandma by themselves even though they don't know the way or to draw strange things all over the road. A realistic road scene will be projected on the side of a building. A citizen trying to divert people around the scene of an accident will improvise a barrier.

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

As soon as our cities are full of these things, the tail will wag the dog. We will start designing our infrastructure around the things they cannot do. We will remove bike lanes. We will reduce people's ability to cross an empty street. Our infrastructure will become more car centric and more Americanized because the Waymos and the Teslas won't be safe otherwise.

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

@gdupont Oh, to be sure. In a way this is nitpicking, but I'm mot fascinated (again, morbidly) not by the bits that show them to have horrifying values, but but the bits that show that they are not nearly as smart as they think they are. It's like all this grandiose language, and then one really idiotic claim like this.

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

And deterrence built on atomic weapons wasn't that sane an idea to begin with.

1

2

0

Peter Bloem @pbloem@sigmoid.social

· 6mo ago

Replying to

@ngaylinn @UlrikeHahn Practically, the models are not as good as we are at a lot of things, but the fact that they are "statistical" is no reason to claim that they are fundamentally incapable of producing truthful answers. Any such argument would apply to people as well. In practice, the story is different. They are imperfect, being manipulated by billionaires and have weird unpredictable failure modes. I totally agree there.

1

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

RE: @paulbiggar@hachyderm.io

I have a horrible sinking feeling that the aftermath of the Iran strikes is going to be the first "but AI said" of international geopolitics.

hachyderm.io

1

4

1

0

Peter Bloem @pbloem@sigmoid.social

· 24mo ago

Replying to

@Ash_Crow @annierau I didn't really get it until 749, then it just clicked.

6

1

0

Peter Bloem @pbloem@sigmoid.social

· 13mo ago

Replying to

@Owlor@meow.social For Cake, a lot would depend on the level of restraint of the audience.

1

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@knowprose Oh, it's not the first time AI has been used horribly irresponsibly. I just wonder if it might be the first time people made the decision to start a war because they blindly trusted some elaborate AI analysis of what the long-term outcome would be.

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

There are people who will tell you that this is all LLMs are, nothing but automated cold reading. That _really_ isn't true. I just wish the AI-kool-aid gang wouldn't make the other side's arguments for them.

0

3

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

I was skeptical about the climate costs of Google pushing a 4Gb AI model to every copy of Chrome, but this article is more careful and reasonable than I expected:

https://www.thatprivacyguy.com/blog/chrome-silent-nano-install/

Specifically:
- It's not the normal "AI inference costs energy" talking point. It's purely about the cost of _transferring_ the model.
- It argues that the problem is the forced mass push.
- It boils down to 240GWh of energy, or 72K homes' yearly energy use.

That Privacy Guy!

Google Chrome silently installs a 4 GB AI model on your device without consent. At a billion-device scale the climate costs are insane. — That Privacy Guy!

Google Chrome is downloading a 4 GB Gemini Nano model onto users

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@ngaylinn This is not new either. All of them going back to ChatGPT have been trained by some form of RL in real-world tasks.

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@abucci @ngaylinn For example, in behaving as a chatbot, under uncertainty does the system hedge over all likely actions (next-token behavior) or pick one and stick with it (agentic behavior).

0

6

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

It's classic security through obscurity. The trick to security is not to keep as many things secret as possible. That way madness lies: you lose track of which secrets are key, and you can't keep an eye on everything. The trick is to keep as few things secret as possible and to guard those carefully. Then make everything else as open as possible. If your system is still safe in that case, then you're secure.

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

If your entire stack is closed source, a hacker can move from one part to the next to get to your source code.
It doesn't have to be Mossad. Any disgruntled employee can run an AI agent over the codebase and sell whatever they find to whoever pays.
Closed source doesn't mean secret source. A codebase is too big to keep secret. It's on every laptop that has it checked out. Snippets are floating around on slack. Shared screens go this way and that over zoom.

0

2

0

Peter Bloem @pbloem@sigmoid.social

· 6mo ago

Replying to

@ngaylinn The innovations here are better finetuning and reasoning traces (that is, having the model output drafts and "reflecting" on them). This is indeed what has reduced the hallucination so much. That's a bit like how you summarize it (hidden prompts), but that doesn't quite capture how rich these reasoning traces can be. See, for example https://arxiv.org/abs/2601.10825 .

arxiv.org

0

13

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

I feel a certain amount of dread knowing that there is a Richard Dawkins novel in progress.

0

Peter Bloem @pbloem@sigmoid.social

· 6mo ago

Replying to

@ngaylinn It's true that they usually don't know "why they know" something (or "know" something) the way we (sometimes) do. It's hard to trace knowledge back to sources. But with RAG this can be done a posteriori. The frontier models with reasoning traces and RAG are pretty good at fact checking themselves.

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@ngaylinn Much of how they are trained is secret but we know that they've been trained on multiple specific tasks since the beginning. In fact this paper first showed that if you do that , they generalize to unseen tasks. This came before the idea of finetuning them to be chat bots. https://arxiv.org/abs/2109.01652 Since then, I think the AI labs have only increased the number of tasks. Things like coding and gaming are good because you can derive rewards automatically, so that's were the focus is.

arXiv.org

Finetuned Language Models Are Zero-Shot Learners

This paper explores a simple method for improving the zero-shot learning abilities of language models. We show that instruction tuning -- finetuning language models on a collection of tasks described via instructions -- substantially improves zero-shot performance on unseen tasks. We take a 137B parameter pretrained language model and instruction-tune it on over 60 NLP tasks verbalized via natural language instruction templates. We evaluate this instruction-tuned model, which we call FLAN, on

0

3

0

Peter Bloem @pbloem@sigmoid.social

· 16mo ago

Replying to

@bert_hubert@eupolicy.social The government usually gets a little more efficient for a few months.

0

Peter Bloem @pbloem@sigmoid.social

· 42mo ago

Replying to

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@ngaylinn This might be a disambiguation issue. What you're describing sounds like the old symbolic RL, where you get proofs and guarantees. What is called RL here is really just a gradient estimator, so you can backpropagate through a sampling process. There are no explicit states or modeling of the problem. Just an action space and a reward signal. I can certainly understand your discomfort. It's completely black box, and people are very rapidly putting it in charge of things.

0

1

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Look at what Mathematicians have been demanding your Respect for ...

0

Peter Bloem @pbloem@sigmoid.social

· 6mo ago

A rare misfire from Tufekci, I'd say. She's usually on the money.

While the algorithm is always the same, AI is operating in substantially different modes when it's hallucinating to when it's retrieving factual information.

Hallucination is not a solved problem, but it's reducing substantially and because concrete innovations, not just "more data, more compute".

0

17

1

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@abucci I don't think we have sufficient common ground to communicate effectively (or at all). I'm going to respectfully end this thread here.

0

Peter Bloem @pbloem@sigmoid.social

· 7mo ago

Replying to

@abucci @ngaylinn That doesn't make it an agent, but it allows for a clear distinction between a statistical prediction model, and one trained towards a specific behavior. Hence, more "agentic". If you don't want to start the discussion before we've resolved the philosophical issue of what an agent is, I don't know what to tell you. I don't think we can afford to wait.

0

Peter Bloem @pbloem@sigmoid.social

· 5mo ago

Replying to

@z_everson@journa.host The guy is president, and has apparently been given a license for open corruption, and he still cannot turn a profit?

0