Peter Bloem

Assistant prof. at the Learning and Reasoning group, Vrije Universiteit Amsterdam.

Not an AI pusher, I promise. I just think we need a better class of criticism.

website: https://peterbloem.nl
github: https://github.com/pbloem
group: https://lr.cs.vu.nl
prons: he/him

Public Key

npub1mjj8dtcld0se03zd42wh9gqjg4mm04ya8azgssagnrvzlvkuq2pqmaht0w

NIP-05 Address

pbloem_at_sigmoid.social@momostr.pink

Profile Code

nprofile1qqsdefrk4u0khcvhc3x648tj5qfy2aah6jwn73yggw5f3kp0ktwq9qspz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dskjwppk

Publishing to

atlas.nostr.land

relay.damus.io

nos.lol

Author Public Key

npub1mjj8dtcld0se03zd42wh9gqjg4mm04ya8azgssagnrvzlvkuq2pqmaht0w

Show more details

Last Notes

2026-05-11 13:00:38 UTC

- reply

Between Google and Apple's insistence on signed software on one side and AI coding on the other, we're soon going to have a serious lack of programmers with any enthusiasm for the craft.

2026-05-11 12:58:50 UTC

I think this will be my last MacBook.
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/556/055/258/000/463/original/c99135ea60854792.png

2026-05-09 12:51:52 UTC

- reply

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

2026-05-06 05:04:17 UTC

- reply

From https://elonlit.com/scrivings/a-theory-of-deep-learning/
Paper here: https://arxiv.org/abs/2605.01172

2026-05-06 05:02:55 UTC

This looks interesting...
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/525/872/854/495/711/original/e12cee1ea8a8a93c.png

2026-04-19 08:59:35 UTC

Mathematicians declare they have solved the crisis. The Strait of Hormuz is, in fact, clopen.

2026-04-18 07:45:17 UTC

- reply

The fact that AI shows clear signs of being in a bubble does not mean that the technology is necessarily on the same level as NFTs and Tulips.
US Railroads, the web, telecoms. All bubbles that burst and did damage, but not because the technology was fundamentally flawed.

2026-04-18 07:42:47 UTC

This seems like a failure of memory to me.
Cellphones were pushed like nobody's business. The polarization around AI reminds me a lot of the attitudes to (pre-smartphone) cell phones. Especially the tedious repetitive arguments.
The telecoms industry was also extremely overvalued and overleveraged in the late 90s.
https://en.wikipedia.org/wiki/Telecoms_crash
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/424/565/420/416/193/original/9a7b42354eca1e11.png

2026-04-13 09:44:23 UTC

RE: https://sigmoid.social/@pbloem/116250532724597189
4 days left to apply.
#note1ly6…lulv

2026-04-09 11:29:20 UTC

- reply

Sorry if saying "do you?" came across as aggressive or combative. It wasn't meant that way. Perhaps I should have said "do we?"
I simply wanted to make the comparison between LLMs and humans. In most of these discussions people paint the LLMs as missing something fundamentally necessary for producing facts. This is usually by contrast with humans.
My point is that there are no _fundamental_ differences.

2026-04-08 12:27:09 UTC

- reply

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 .

2026-04-08 12:22:06 UTC

- reply

The "statistically probable" part refers to the pre-trained model, not the fine-tuned one.
One example is asking "when's my birthday". The pre-trained model doesn't know, and so picks a random answer from all possible dates. That's hallucination.
The finetuned model takes that uncertainty, but expresses it as the answer "I don't know".
Sometimes the finetuning fails and we get hallucination in chatbots. This is because they are overtrained to be certain.

2026-04-08 09:28:51 UTC

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".
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/368/358/626/949/505/original/78aceb5a020d9f73.png

2026-03-18 13:59:17 UTC

Do you feel like AI could learn to reason a bit better? Do you have ideas about how to make this happen? Do you feel like AI does reason, but we have no idea how and we should study it?
Well, we at the Vrije Universiteit Amsterdam are hiring an assistant professor to study just these kinds of questions. The profile is "learning and reasoning" in the broadest sense. If you can explain how your research touches these two topics, we want to hear about it.
https://werkenbij.vu.nl/vacatures/assistant-professor-learning-and-reasoning-career-track-amsterdam-1278904
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/250/538/875/307/211/original/5c499fea0e11fcfe.png

2026-02-26 08:36:55 UTC

- reply

David MacKay discusses this in Sustainable Energy without the Hot Air, and it's pretty disappointing (for fundamental rather than technical reasons). The UK can get at most 1.6% of its energy budget from geothermal.
If you divide the amount of heat the earth dissipates by the number of people, the resulting number is simply pretty low.
It works for places like Iceland, because a large area's worth of heat comes out at one point where the population is low.
https://www.withouthotair.com/c16/page_96.shtml

2026-02-22 11:59:16 UTC

- reply

Also Mastodon/Fedi? I mean AI is pretty unwelcome, but the other disciplines seem to be having a good time here.

2026-02-22 11:40:08 UTC

- reply

I won't do it. I won't become a LinkedIn guy.

2026-02-22 11:38:45 UTC

It looks like the AI research community is finally moving away from X and of all the possible platforms, they've picked LinkedIn. These fucking people...

2026-02-14 11:30:24 UTC

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.

2026-01-27 08:56:53 UTC

The thing about vibe coding is, I was never really on board with my IDE closing my brackets for me.

2026-01-04 13:19:41 UTC

RE: https://mastodon.social/@nixCraft/115835690251372850
This is, of course, a problem for AI as well. StackExchange is a hugely important part of the pre-training data. Maybe even more important than Wikipedia.
For anything that happened after ~2024, that pre-training data is now no longer available.
Models have got pretty good at working novel problems in-context, but not nearly as good as they are for in-distribution problems (stuff that SE had questions on). We'll have to see if AI keeps working as well on really new topics.
#note1jy0…xyua

2025-11-28 15:10:03 UTC

- reply

In a way all that pre-training power is just initialization. Our initial brain state is also the result of a huge power investment.
What matters is how well they learn in context.
Obviously, an M4 running Llama 13B is definitely nowhere near human intelligence. Still, it doesn't seem at all obvious to me that we would need completely new hardware to get to human intelligence @ 20 Watts.

2025-11-28 15:06:56 UTC

I made the familiar point yesterday during a panel discussion that even if AI were approaching human intelligence, we're nowhere near human power efficiency (in "intelligence per watt")
But is that actually true anymore? We need huge power to train models, but once we have them trained, distilled and quantized, what do we have?
The human brain draws abt 20 Watts. An M4 doing local inference on a heavily quantized LLM won't draw more than 40 Watts.

2025-06-29 13:23:33 UTC

Solving ARC (well... up to 20%) with inference-time compression. Another indication that #MDL and #AlgorithmicComplexity are becoming relevant in the age of Deep Learning.
https://iliao2345.github.io/blog_posts/arc_agi_without_pretraining/arc_agi_without_pretraining.html

2025-06-11 12:54:11 UTC

I asked ChatGPT to transport me to the airport and it couldn't do it. My 30 year old car gets me there just fine and uses much less energy. This technology is useless.

2025-06-09 21:41:44 UTC

If you want to help people in #academia who are maybe less fortunate than you, who have less famous supervisors, or work at less prestigious universities, here's one simple thing you can do:
Do proper literature research.
That means complete forward and backward snowballing from a decent seed set. Find everything that is relevant to your paper and cite it. Budget a couple of full days for it.

2025-06-06 09:51:57 UTC

An 8TB corpus of copyright-free text for training AI models. Comes with two 7B models trained as proof-of-concept.
https://github.com/r-three/common-pile/blob/main/paper.pdf
I'm glad somebody has finally done this. The "we need to break copyright or AI won't work argument" feels super-dodgy to me and we can just evaluate whether it's true. 1/n
https://cdn.masto.host/sigmoidsocial/media_attachments/files/114/635/791/207/485/953/original/a0ffca5dad2751ad.png

2025-03-09 18:38:59 UTC

New lecture on transformers for my #MachineLearning BSc course.
We go all the way from simple self-attention to a full-fledged ChatGPT. Includes position embedding, tokenization, instruction tuning and much more.
https://mlvu.github.io/transformers/
https://cdn.masto.host/sigmoidsocial/media_attachments/files/114/133/880/525/589/457/original/3f2c47abba1b9af2.png
https://cdn.masto.host/sigmoidsocial/media_attachments/files/114/133/883/031/379/855/original/24dc0cbee0ddeba2.png
https://cdn.masto.host/sigmoidsocial/media_attachments/files/114/133/887/216/727/579/original/531c183f24fa5e75.png
https://cdn.masto.host/sigmoidsocial/media_attachments/files/114/133/893/158/827/809/original/8941de5b870768ec.png

2025-02-28 14:39:36 UTC

- reply

Is the cute cartoon a dark pattern? I like cute things (and this fox is well-drawn to be fair).
But I think it's a deliberate attempt to infantilize the user and to put a harmless face on something insidious.
@npub1p09…m5vl noted this more than ten years ago. Make big brother a cartoon dinosaur instead of a mustachioed man and suddenly we will swallow anything.
https://ar.al/notes/spyware-2.0/

2025-02-28 14:35:44 UTC

Moving away from Firefox shows how much the rot has already set in. After dismissing the "set as default browser" dialog, and explicitly clicking "don't ask again", next time it shows me a different pop-up, that is much easier to accidentally click.
All the dark patterns are there: high contrast easily-clickable default option. "No now" instead of "No" as low contrast alternative answer. Cute little cartoon.
https://cdn.masto.host/sigmoidsocial/media_attachments/files/114/082/001/025/522/551/original/e7a32efc004245a2.png

2024-12-23 10:22:11 UTC

I cannot express my frustration with Intel and AMD at this point. It's 10 years and counting since deep learning should have been on everybody's radar and they still cannot seem to grasp that you need _software support_ to make people use your chips.
This NVidia monopoly on #ai hardware is not good for anybody. We need a little competition in this space.
https://semianalysis.com/2024/12/22/mi300x-vs-h100-vs-h200-benchmark-part-1-training/
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/701/636/904/587/818/original/3eb055d557aa06ca.jpg

2024-12-16 18:29:22 UTC

- reply

If you get this sort of shit, I made a handy website a while ago.
https://dontmail.it

2024-12-16 18:28:41 UTC

my insurance company: Don't fall for phishing mails, always check that the sender is us. That's how you can tell who emails are from!
also my insurance company: Time to pay the bill, here's a handy link to some third party company you've never heard of.

2024-12-16 16:18:20 UTC

- reply

Why is this not quoted more among computer scientists and programmers? Why is it not played on a loop throughout the year, from speakers around campus?

2024-12-15 15:42:41 UTC

- reply

Even that didn't really split the fediverse, I think. I'm pretty sure I'm still talking to people on either side of the divide.
I think that's the key. Because there are so many instances, there will always be little bridges between any two sides of a split.
We never really talk to all of the fediverse, but any message has a chance to reach a good chunk of it.

2024-12-15 14:51:33 UTC

- reply

I think it probably wouldn't. Unless people actively defederated from instances that would defederate over hosting a wedge user. But if the aim is to defend freedom of speech, that wouldn't make much sense.
The result is probably that the wedge user gets their platform on an instance somewhere, but only reaches a limited slice of the fediverse.
It's all a bit messy, but I do think the fediverse is more robust to this sort of thing than a centralized platform with a single policy.

2024-12-15 14:47:43 UTC

Bluesky has been shaken up by the joining of somebody who half the people strongly feel should be banned and half the people strongly rally behind. A kind of wedge-user.
Without discussing the specifics, I wonder if the #fediverse is more resilient to this. If a wedge-user joined a large instance, and they were banned (which would prob happen in this specific case), they would then move to an instance that would take them.
Would this cause a split into effectively two fediverses?

2024-12-09 18:46:50 UTC

- reply

This quote is showing itself to be more and more prescient.

2024-12-08 14:14:12 UTC

If you are an #artist worried about #ai being used to copy your style, it is a miscalculation to worry only about training data.
Many artists are explicitly forbidding use of their work in training generative AI models.
However, a large amount of research is looking into in-context use of examples. For example, you give ChatGPT three examples of poems and then ask it to generate a new poem in the same style.
(1/6)
Rest of thread: https://sigmoid.social/@pbloem/113617628089668394
#artistsofmastodon

2024-12-01 15:34:46 UTC

- reply

I thought it was a kind of reluctance to collective action and general inertia, and taking the road of least resistance.
I'm now beginning to think most of them are fully on board with the whole enterprise.

2024-12-01 15:33:30 UTC

- reply

It's true, that it's surprising. However, as an AI researcher, I'm slowly beginning to learn how naive I've been about my community
I thought they were all basically scientists raised on Star Trek. So how bad could it be? Then they all went to work for Google, OpenAI and FB. Then OpenAI went private and they all kept working there. Then Musk took over X and they all stayed there. Then X put Trump on the throne and they're all still there, some of them applauding.

2024-12-01 12:28:12 UTC

- reply

I don't know about FB, but during the Twitter takeover, it was pretty clear that Musk gutted the company of anybody who wasn't absolutely dependent on job security for their VISA.
Everybody who didn't leave in disgust was going to do whatever he said with no questions asked.

2024-10-23 07:24:00 UTC

https://onesquareminesweeper.com/

2024-10-21 19:00:44 UTC

I wrote a new article.
⬬ Understanding Gaussians ⬬
https://gestalt.ink/gaussians
The idea is that Gaussians can seem pretty daunting, especially if you look at the density function. However, they're actually simple to understand if you forget about all that and focus on _linear transformations of the standard Gaussian_
Then, the density function gets really simple, and the rest follows from basic rules of linear algebra.
#tutorial #machinelearning #statistics #explainer #math
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/346/925/314/803/566/original/f8c68c838bd1e81d.png
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/346/932/831/494/917/original/e975770b028d6472.png
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/346/941/209/180/552/original/7b22155568b5c55a.png
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/346/945/711/909/762/original/0642de0e3126053c.png

2024-10-14 08:04:46 UTC

Assuming you're an average meat eater, you take two intercontinental flights per year, you drive a car daily, and use ChatGPT/Copilot regularly throughout the day.
Which of these activities should you cut to have the biggest positive climate impact?

2024-09-22 16:48:20 UTC

- reply

But would either of these lower the ecological footprint? I think the models would just grow. How much energy and money is spent is currently determined by the companies' future expected reward, not by environmental concerns or hardware limitations.
And the expected reward is probably inflated by the arms race and them drinking their own Kool aid.

2024-09-20 13:10:49 UTC

I agree that #AI has an environmental problem, but Mastodon's default position of "anything to do with AI is inherently evil" does not feel particularly productive.
I find a lot of AI plenty useful, even LLMs. If I'm going to take the environmental costs seriously as a user, I need to talk about it much more precisely, and carefully than what's happening now.
Saying "AI is an environmental crime of gigantic proportions." is so over-the-top, there's no meaningful conversation left to have.

2024-08-17 10:19:27 UTC

- reply

I tried Obsidian, but it made me realize that all I need is a dropbox directory with markdown files.
If I'm on my laptop I can edit them in an IDE or text editor, and on my phone I use markor & dropsync.
I get no lock-in, all the formatting offered by markdown and search through spotlight. If I ever need more organization, I can fire up Obsidian again and it'll load it all.