Natürlich zitiere ich Helmut Schmidt bei einem Text über den anstehenden Monster-Börsengang von SpaceX. Was könnte man sonst machen, wenn man aus Deutschland auf so ein Unternehmen schaut?

I can imagine finetunes of this thing going insane. Mogs OSS-20B on agentic stuff. I assume it's really fucking fast because of how few active params it has. 38T tokens (more than deepseek-v4!!!) for an 8B model wew

paradoxically this is because anthropic crushed them so thoroughly in the code-agent market that they demonstrated that ai is potentially extremely profitable, and being second or third is also so valuable that this will outrun oai's debts

*and* they are profitable

Social media has been a real window into the amount of people whose politics basically works backwards from "But *I* want to be in charge of the purges".

Two things can be true simultaneously: a) Modern LLMs can count the amount of letters in a word despite the counterintuition of tokenization b) Google Search Overview's LLM can fail to the amount of letters in a word because it's a quantized LLM There's a nuance that tbh no one care about anymore.

MiniMax AI consolidated all of the work behind M2 and published it. Paper: arxiv.org/abs/2605.26494

In 50 years discourse will talk about how LLMs were the good type of automation and this new evil type (scalable molecular printing or whatever idk) is bad and horrible and not like LLMs at all.

and models use ~15 mL each if you include the offsite costs, roughly 5 mL in the data center itself." It's silly that in 2026 educated people still believe that AI uses a bottle of water per normal prompt. No one investigating this believes it anymore. (5/5)

I've had a very nice back-and-forth email exchange with Shaolei Ren, who provided the original "AI uses a bottle of water per prompt" estimate with the Washington Post. (1/5)

People don't long for "inconvenience." They long for permanence. And though streaming services may be convenient, people are waking up to the fact that we are paying rent-seeking middle men month after month for things we will never actually own.

there’s so much overhang that they didn’t even think to try the Erdos problem

It was just 8 months ago when people made fun of Altman who talked about PhD level intelligence in domains like math. Now we have Fields Medal winners talking about LLMs doing PhD-level research. And expressing concerns about math as a career.

I recently, together with my eldest, got into painting warhammer figures. I used Dia to summarize yt tutorials and save those that to notion, I then used Claude to compile lists of the most used/recd colors & painting tips. etc. pp. AI is your best companion to help you get into a new topic.

GOOD NEWS! Researchers at Stanford University have developed a UNIVERSAL vaccine known as GLA-3M-052-LS+OVA, that protects against a wide range of respiratory viruses, bacteria AND even allergens. The vaccine is delivered intranasally AND provides broad protection in the lungs for several MONTHS.

"why would i read something i was going to criticize" say only the most intellectually honest people: the ones who already know the right answer.

Nine more Erdos problems resolved with LLMs + Lean: arxiv.org/pdf/2605.227... Worth noting they got 9/353 after trying a giant set of open problems by Erdos, and 44/492 OEIS open conjectures. Likely would've gotten 0 a year ago.

Ich habe vielleicht ein Problem. Und wie so oft kann mir nur KI dabei helfen. (Dia hat den Wust gesichtet und sortiert bevor ich die Hände in die Luft geworfen und "alles verbrennen" gerufen habe.)

it's just so weird to be continuing to call this tech "spicy autocorrect" as it solves math problems no human has we have AI. actual for real AI. you might have problems with how it's being distributed or trained, but it *works*

Hey guys, don't worry about AI solving an Erdős problem, at least you can all rest assured that the output cannot be copyrighted.

Tipp für alle mit Agenten und lokalen Dateien (also für alle mit Computer): Lasst euch von Claude Code/Cowork oder Codex oder Hermes, Openclaw etc. pp. ein regelmäßig aktualisiertes HTML-Dashboard bauen, das ihr im Browser als Tab anpinnen könnt. Einfach umsetzbar, anpassbar & extrem praktisch.

Banger thread

One thing I've kind of expected for awhile is a lot of longevity problems get solved in cats & dogs first because the regulatory barriers are lower, though hopefully a lot of that research will carry over to humans

Haben einen ausführlichen Faktencheck in meine Schreibapp eingebaut. (von Claude einbauen lassen) 🥳

Schaeffler hat jüngst gleich mit 3 Roboter-Startups (Humanoid, Neura Robotics, Hexagon) interessante Deals abgeschlossen. Sie werden zum Lieferanten für Aktuatoren (Motoren für die Gelenke) und zum Abnehmer für die humanoiden Roboter der Startups, die dann in den Werken von Schaeffler arbeiten.

haha

heh

There has been a lot of speculation that AI companies were unprofitable, but Anthropic will have an operating profit of $559M this quarter. “In the first quarter, Anthropic spent 71 cents on computing power for every dollar it made. In the current quarter, it expects to spend 56 cents per dollar…”

OpenAI's claim that this is a central conjecture in discrete geometry is not an exaggeration. This will I think be looked back on as the first time that AI solved a major mathematics problem (defined as a problem that all experts in some subfield had thought about). openai.com/index/model-...

Woah!

We can estimate the resource cost of solving the Erdos problem. The calculations below seem reasonable, so using the best public estimates we have, it took 0.6–6.3 kWh of electricity & 3–31L of water Equivalent to < 3 almonds worth of water & the electricity equivalent of 2-20 miles of EV driving

unfortunately openai didn't publish the unsummarized chain of thought, but the summary is 125 pages! the model reaches the crucial idea (which it describes as 'frightening,' i would love to read the unabridged chain of thought here...) on page 39

June 2024: The latest general-purpose LLMs could not count the r's in strawberry. July 2025: The latest general-purpose LLMs get gold in the International Math Olympiad. May 2026: The latest general-purpose LLM solve an 80 year old problem, one of the "best-known questions in combinatorial geometry"

JFC! benthams.substack.com/p/nestle-ki...

I can't believe we turned "Laptop Worker" into a term of derision. We now have the tech and business models to allow millions of people to make a living in ways unthered to a farm plot, assembly line,, or cubicle. That's very cool! Do not project your weird yearning for the mines on to me.

People who see value in AI are explaining popular anxiety by saying "SV did a bad sales job." Possible. More plausible, for me, is @michellegoldberg.bsky.social: anxiety is highest in the US b/c low trust in leaders generally. But also, come on—why would we ever expect AI not to provoke anxiety?

“friction is good actually” is such a stupid argument, we’ve been removing friction from every system since the dawn of man but now that you’re not in your 20s anymore and you’re used to how the world works everything that sucks is now good and we have to do it that way forever unchanged

Hats off to the AI2 team. This is data that is openly available, but put behind paywalls elsewhere as "proprietary." AI2 is making data that once cost thousands available at a rate that's affordable to everyone. This is GOOD AI disruption, because EVERYONE who has a use for this data can now benefit

Tag 7, 197.000+ Pakete gescannt. Die Roboter von Figure arbeiten 24/7. Die PR-Expert:innen bei Figure sind echt gut. www.youtube.com/watch?v=luU...

AI agents one-shotting apps and small personal everyday use tools for non-programmers *could* lead to the year of the Linux desktop. 🙃

I have written a new post, "Implications Of Predicting The Next Token" in which I explain why saying a language model "just" predicts the next token fails to understand what that phrase would even imply. minihf.com/posts/2026-0...

Der selbe Prompt, ein Jahr Unterschied: (Quelle: www.linkedin.com/posts/peter... )

Should you have ChatGPT write all your emails? Probably not! But to some people, writing that email is an insurmountable challenge, and having the robot write them up a draft (or even the whole thing) can help them overcome it. "It's just an email, what's the big deal" is literally "smile more".

"Remember, you were perfectly capable of writing that email before LLMs" is a performative, anti-science, and worst of all: ableist take (no, not everyone was perfectly capable of writing that email), and to see it wielded by folks who previously made fighting ableist language their brand is telling

Europas Autokonzerne werden nicht (nur?) frontal von chinesischen Herstellern verdrängt, sondern von zwei Seiten zerrieben: oben von Mobilitätsplattformen wie Uber und Waymo, unten von chinesischen Technologie- und Rohstoffstacks bei Batterien, Software und Halbleitern.

All along the watchtower

I know people on Bluesky hate AI and I am opening myself up to having digital tomatoes thrown at me here... but digital tomatoes don't stain clothes, so here goes nothing: How can anyone honestly think that being able to interact with computers in natural language *won't* be a big deal?