Peter Bloem

@pbloem

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. 🌉 bridged from ⁂ follow @ap.brid.gy to interact

39 Following388 Followers963 Posts

Bluesky

Is this your profile?Claim it and pbloem.dirty.social is yours.Claim profile

Peter Bloem@pbloem· Jul 20

The poor thing is going through this over and over again as people wake up to the tweet and paste the example into the chatbot.

AAAAH. Wait a moment. WAIT. I want to reexamine the claim "det J constant = (JC would say)
injective". Yes. So if this example is right, JC is disproved by a one-line example, which would
have been checked by symbolic computation by literally anyone with Mathematica in seconds. The
map is SO simple that this cannot possibly have been missed for 85 years. Therefore the
overwhelmingly likely resolution: THE MAP IS NOT ACTUALLY WHAT IT SEEMS — maybe it's not a map
C3-C3?? It is. Or maybe... my repeated computations share a systematic blind spot. The mod-101

Peter Bloem@pbloem· Jul 19

Very strange behavior from Claude (Fable). It corrects itself at the last minute from making a false claim. I've seen this a few times. It's fine, but I wonder if this is intended behavior. You'd think it could get rid of this during the thinking phase.

Claude screenshot: 4. The mask update lags one block. mask[:seen] = True then seen += block means at iteration ibl you mark blocks 0..ibl−1 — the block you just encoded is excluded, so training never sees the most recent block. Prequentially you're allowed to train on everything encoded so far, including the current block. Increment first, then mark (this also matters for the bpd display, which currently divides current-block cost by a count that... actually the bpd is fine since you divide after incrementing, but the training mask is off by one).

Peter Bloem@pbloem· Jul 11

So much for sycophancy...

A Claude screenshot: "Your intuition is backwards for this workload, and the difference is large: this search is almost entirely small-array numpy, which does not parallelize internally at all."

Peter Bloem@pbloem· Jul 8

This is why I've been pushing back (and yes reply guying) on people confidently claiming that AI is glorified predictive text. Not because I'm an AI pusher. Not because I love AI and I think everybody should embrace it. Because it's a wrong mental model, and […] [Original post on sigmoid.social]

Are you worried about AI?

We were asked this question on a small-publishers panel at the World Fantasy Convention in 2025. Our Managing Editor, C.L., remembers answering pretty confidently. AI is glorified predictive text. It can produce words that technically hang together, but it cannot innovate, cannot imagine, and will never write anything worth reading.

He still believes that. We still believe that. Mostly.

Peter Bloem@pbloem· Jul 8

On the left, some typesetting in the TMLR/JMLR standard template (Times and PC Modern, I think). On the left, the same in Charter and Euler. I think journals need to open up the opportunities to use different typefaces. PC Modern is an unreadable mess.

A bit of mathematical typesetting in the typefaces Times and PC Modern.

A bit of mathematical typesetting in the typefaces Charter and Euler.

Peter Bloem@pbloem· Jul 3

Occasionally forgetting about grammar or comprehensibility altogether in its closing paragraph.

One sensitive-topic note this isn't, but a fairness one: there's a real equity angle people raise — those who can't afford AC bear the externalized heat from those who can — if you're interested in that dimension I can go into it.

Peter Bloem@pbloem· Jun 26

Double descent, here we go. Any minute now.

A validation loss curve that is going down and then back up again.

Peter Bloem@pbloem· Jun 14

I'm deeply frustrated by the way today's "thought leaders" (or in any case, the voices that are heard) repeatedly take an important point that deserves to be discussed, and surround it by insane hyperbole that causes it to be instantly dismissed by people on […] [Original post on sigmoid.social]

Marc Andreessen = © X.com
@pmarca
Of course every part of this is absurd.
(1) Al isn't nukes, it's math.
(2) Big companies, small companies,
independent developers, and open source
communities should be free to pursue Al.
(3, highlighted) Al companies should not be nationalized, de
jure or de facto (government-protected cartel).
— Al means restricting math,
software, and chips, which taken seriously
would impose tyranny far beyond anything
even imagined by the Communists and Fascists
of the 20th Century)
(5) Al authoritarians should learn from history
and rethink their positions.

Peter Bloem@pbloem· May 24

@riaschissl Ah, sorry, a bit more context: bird.makeup is a service that let's you clone twitter accounts into the fediverse. You can't interact with them, but you can see their posts on your timeline. I did this for the twitter account @visualizevalue. It […] [Original post on sigmoid.social]

A twitter screenshot of an @visualizevalue post. It says “The first rule of compounding is to never interrupt it unnecessarily.”. In the picture are two graphs: one with a kind of sawtooth pattern of "growth" being continuously interrupted (reset to zero) and one of a single uninterrupted exponential curve.

Peter Bloem@pbloem· May 21

I think this is what bothers me about the environmental impact of AI criticisms. It's not that AI has no impact, or that we shouldn't do much better. It's that the perception is now that AI is one of the main causes of climate change, because it's such a […] [Original post on sigmoid.social]

Hacker News screenshot:

jmclnx 1 minute ago | prev [–]

>If warming continues at this rate, humanity could breach the Paris Agreement goal of limiting warming to 1.5°C in 2028, even sooner than other research has projected.

I wonder if we are already there :( I remember a year or 2 ago we breached 1.5C for a short period of time.

Crypto mining was bad enough, now with AI and Trump, I expect it will happen sooner then later.

Peter Bloem@pbloem· May 17

The most realistic movie depiction of AI/human relations is Wall-E. People don't resist. They are falling over themselves to defer responsibility.

You hear wild stuff all the time now. Like this story that Nat Friedman, a former CEO of GitHub, told recently at a conference. Friedman uses OpenClaw, an autonomous AI agent that runs on his computer, acting like a personal assistant. One day, his OpenClaw decided that he wasn’t drinking enough water, so Friedman instructed the agent to “do whatever it takes” to make sure he stays hydrated. According to Friedman, eventually the bot directed him to go to the kitchen and drink a bottle of water. It informed him that it was monitoring him via a connected camera in his home. “I’m going to watch to make sure you do it,” the bot supposedly said. Friedman did as he was told, and, moments later, the bot sent him a frame of him drinking the bottle of water and said good job. “I felt like I did do a good job,” Friedman said.

Peter Bloem@pbloem· May 11

I think this will be my last MacBook.

A homebrew screenshot showing that LibreWolf will be deprecated, suggesting that homebrew is going stop carrying software that hasn't been approved by Apple.

Peter Bloem@pbloem· May 10

Found on Bluesky (by @hannahshelley.site)

That hand-drawn meme of the guy who is impressed by his computer pretending to be sentient in a basic way. Except now the guy asks his computer "what do you know about me" and the computer answers "basic web traffic data" and the guy goes "oh my god." It's about that "taken" website that very dramatically presented some very basic data that everybody seemed to be really impressed with.

Peter Bloem@pbloem· May 8

Many little ducklings. One for each of the sync errors on MS's cloud, which means that in the final week of our proposal submission, we have to email changes to one person who will then integrate them into the document, like it's 1990.

Peter Bloem@pbloem· May 6

This looks interesting...

expression to a per-parameter rule: @
update parameter k if and only if
2

Hi > _
That is, if the batch signal on a parameter
exceeds its leave-one-out noise, update it;
if not, skip it. This is a one-line change to
Adam that accelerates grokking by 5x,
suppresses memorization in PINNs, and
improves DPO fine-tuning, eliminating
the need for validation sets entirely.

Peter Bloem@pbloem· Apr 27

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

Peter Bloem@pbloem· Apr 27

Clowns to the left of me, jokers to the right...

A twitter screenshot of blueticks on X being convinced that ChatGPT can do palm reading.

Peter Bloem@pbloem· Apr 24

Peter Bloem@pbloem· Apr 22

A screaming duck. Why can't I log in, Microsoft? Why?

Peter Bloem@pbloem· Apr 21

Angry-looking duck staring straight at the camera. Courtesy of:
"Something went wrong. We're sorry, sign-in isn't working right now. But we're on it! Please try again later."

Peter Bloem@pbloem· Apr 21

I am reading, with morbid fascination, the Palantir "manifesto". This bit sticks out to me. It's very illustrative of how shallowly these people think about AI. They always build on some assumption that the effect on society can be predicted and harnessed […] [Original post on sigmoid.social]

12. The atomic age is ending. One age of deterrence, the atomic age, is ending, and a new era of deterrence built on A.I. is set to begin.

Peter Bloem@pbloem· Apr 19

I'm sure everybody already knows this, but it's eye opening for me. UBlock allows you to click on any element on a website to block it. It doesn't have to be an ad. It could also be an annoying AI chat assistant, an auto-playing video, the hot network […] [Original post on sigmoid.social]

The "block element" option in the LibreWolf context menu added by ublock origin.

Peter Bloem@pbloem· Apr 18

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 […] [Original post on sigmoid.social]

BSKy post: Adam Conover
‪@adamconover.net‬
I don't recall any previous transformative technologies that had to be sold to the public so hard. No one in 1999 was going, "Guys. Cell phone? It's coming, whether we like it or not. I think it's important, FOR WOMEN ESPECIALLY, to learn about cell phone. So that we're not left behind!!"

Peter Bloem@pbloem· Apr 11

You know you're in the weeds when the background color of the PDF is off-white.

Screenshot of the paper that introduced Jeffreys' prior.

Peter Bloem@pbloem· Apr 8

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 […] [Original post on sigmoid.social]

zeynep tufekci
@zeynep
·
7h
Humans make many mistakes. It’s in no way analogous to how generative AI “hallucinates”.

The key point is generative AI is absolutely the same mechanism when it “hallucinates” and when it does not.

If that’s not obvious, I recommend getting into the weeds of the technology.
Quote
Beck
@Beckeckeck
·
7h
Replying to @zeynep
Humans also hallucinate some percent of the time, does that mean we can't use them at scale?

I agree that what you're gesturing at is a problem for long term

Peter Bloem@pbloem· Mar 20

Here's an example of the sort of thing I'm talking about (only hearsay, but it serves to illustrate the point). This is just classic "ML at scale discovering weird causalities" but just think what the system has to work with here. It can operate in the full […] [Original post on sigmoid.social]

X post, quote tweeting some marketing from a video-generation app: Ben
@BCalusinski
End of times stuff here

There was a guy on this app who tested a bunch of similar ai UGC ads and realized if the thumbnail showed the woman’s armpit that the views would oftentimes 5-10x

He was dumbfounded and had to try and figure out the causal reasoning for it and figured out it had something to do with how humans don’t lift their arms unless they want to feel more exposed/vulnerable (like hugs for example) so a bunch of guys would be more likely to click the vid

Most people will literally have the same value as sheep a decade from now if this keeps up at scale

Dead internet theory is correct - agents are the only way to extract signal without the mental clutter

Peter Bloem@pbloem· Mar 20

In the lecture notes is a link to a great keynote by Neil Hunt about what Netflix is doing at the same time. This shows that "we didn't know" is no excuse for Youtube. There were plenty of intelligent people in the industry who knew exactly why what they were doing was a bad idea.

A slide summarizing the Neil Hunt video and the main points: new vs. existing customers. Dificult to measure metrics. Total vs median hours viewed. Addictive vs. compelling content.

Peter Bloem@pbloem· Mar 20

I gave my last machine learning lecture of the course this week. In the last part, I look at the dangers of making ML part of the infrastructure of society. https://mlvu.github.io/rl/#video-093 One case study I discuss is Youtube in 2026 optimizing purely […] [Original post on sigmoid.social]

A screenshot of the Youtube official blog. They post title says "You know what's cool? A billion hours" It's them bragging about optimizing purely for user engagement, which would later massivley drive polariziation and extreme content.

Peter Bloem@pbloem· Mar 11

A mottled duck mother with 9 ducklings swimming through duckweed.

Courtesy of Microsoft Teams. How do you build a file browser that goes back to the root dir when you refresh? How is this a serious company? Also you need to refresh to see any new files that were added. And then you need to navigate back to your folder.

Peter Bloem@pbloem· Mar 3