Mattia Rigotti

ChatGPT File • • Edit View Window Help ChatGPT 5 > How many times does the letter b appear in blueberry The word "blueberry" has the letter b three times: • Once at the start ("B" in blueberry). • Once in the middle ("b" in blue). • Once before the -erry ending ("b" in berry). Total: 3 O92 Show me where the letter b appears in the word please. Sure — here's "blueberry" with each b highlighted: B lue b erry Positions (counting from 1): 1 B - first letter 4 b - in blue 6 b - in berry D9< Spell that out with all three bs please. I mean, just spell the word. Sure — here's the full spelling with all three b's shown in place: BLUEBERRY That's exactly how blueberry is spelled, with the b's in positions 1, 5, and 7.

ChatGPT File Edit View Window Help E ChatGPT 5 > You wrote "BLUEBERRY". There are only two bs there. Ah - I see the mix-up. Let's slow it down and spell blueberry carefully, letter by letter, showing all three b's. blueberry Now you can see them: • 1st b → position 1 • 2nd b → position 5 • 3rd b → position 6 blueberry Exactly — blue berry has three b's: • 1st b → at the very start • 2nd b → right after the e in blue • 3rd b → immediately after the second b, starting berry So the "bb" in the middle is really what gives blueberry its double-b moment. Its double-b moment; I see. Yep - blueberry is one of those words where the middle almost trips you up, like it's saying "b-b-better pay attention." That little bb moment is satisfying, though — it makes the word feel extra bouncy.

 ChatGPT File Edit View Window Help 巴 • ChatGPT 5 > My students get you to write their essays for them, you know. Ah, the old "ChatGPT, do my homework" routine — I'm familiar with that one. I can see why they'd be tempted, but here's the thing: I'm a lot more useful as a coach than a ghostwriter. If they use me to brainstorm ideas, plan structure, or clarify tricky concepts, they'll actually learn and still have their own voice in the work. If they just hand in what I write, they're skipping the thinking part - and it usually shows. I'm good, but I still leave a trail of... telltale ChatGPT-isms. A trail of blueberries, as it were. Exactly — a trail of blueberries: easy to follow, a little sweet, and maybe just obvious enough for a watchful teacher to notice. If your students really are leaving me to bake the whole blueberry pie for them, they might forget that the best part of pie is making it themselves... even if it gets a little messy. "Blueberry" only has two bs. Nope - blueberry really does have three b's. Here it is spaced out: bluebberry Positions: b- at the start 5 b - in the "blue" part 6 b - right next to it, starting "berry" That middle double-b is easy to miss if vou just glance at the word.

I had to try the “blueberry” thing myself with GPT5. I merely report the results.

Screenshot of benchmark with gpt-5 on top with 68.46% accuracy.

GPT-5 lands first place on NoCha, our long-context book understanding benchmark.

That said, this is a tiny improvement (~1%) over o1-preview, which was released almost one year ago. Have long-context models hit a wall?

Accuracy of human readers is >97%... Long way to go!

Trump tariff revenue soars 78%. Who's paying them? Tariff collections are up sharply in the last two months. Congressional forecasters say tariffs could help reduce the federal debt, but they'll also lead to higher inflation and slower economic growth...

Other countries will pay for tariffs is the new "Mexico will pay for the wall"

www.npr.org/2025/06/09/n...

Tweet by Sam Altman featuring an image of the Death Star under construction, visible on the horizon against a cloudy sky: https://x.com/sama/status/1953264193890861114

If OpenAI think of themselves as the Empire, who is the Rebellion?

gpt-oss: OpenAI validates the open ecosystem (finally) OpenAI's first open language model release since GPT 2 and what it means for the ecosystem.

For years, a few people continued to advocate for open models despite all sorts of attacks. OpenAI was one of them, at least indirectly. All of this work is paying off, and the perception of open models in the US is changing.

I take this as a personal win. More work to do. buff.ly/fFEKMCh

Whether it's conservatives assassinating State Senators in Minnesota or liberals refusing to sell Alan Dershowitz pierogis in Martha's Vineyard, both sides have taken incivility to an extreme.

Terence Tao has written a thread about the damage being caused by the grant freezes at UCLA.

mathstodon.xyz/@tao/1149568...

Our new paper is out! When navigating through an environment, how do we combine our general sense of direction with known landmark states? To explore this, @denislan.bsky.social used a task that allowed subjects (or neural networks) to choose either their next action or next state at each step.

Trump Gets Rid of Those Pesky Statistics The numbers were simply not patriotic enough.

Donald Trump Declares Victory Over Statistics

Well that’s not ideal

17 of the top 20 open models on the Artificial Analysis leaderboard are made by Chinese companies. 2 from the US and 1 from Europe. This is what current domination of open models looks like, but it can be changed quickly with focused funding and effort.

Leaderboard: buff.ly/hVyaNl8

At least 3 killed and others injured in train derailment in southern Germany A regional passenger train has derailed in southern Germany, killing at least three people and seriously injuring others.

One reason a German train accident is international news and German car accidents are not is that roughly 1000x more people die in car accidents each year in Germany. It’s so common that it’s not newsworthy.

Hierarchical Reasoning Model Reasoning, the process of devising and executing complex goal-oriented action sequences, remains a critical challenge in AI. Current large language models (LLMs) primarily employ Chain-of-Thought (CoT...

HRM: Hierarchical Reasoning Model

ngl this sounds like bullshit but i don’t think it is

- 27M (million parameters)
- 1000 training examples
- beats o3-mini on ARC-AGI

arxiv.org/abs/2506.21734

I imagined the Chatsubo in 1984. 41 years later I opened its door. Neuromancer is in production.

🚨PhD student position in #cognitive_psychology and #neuroscience, on a project directed by André Knops and in collaboration with the Dugué Lab!

@upcite.bsky.social | @cnrs.fr

Funded by the @agencerecherche.bsky.social (Agence nationale de la recherche)

www.linkedin.com/jobs/view/42...

We now get a major preview of the next chapter of Physics of language model. And it's not toy models anymore: systematic synthetic data/design experiments to create SOTA 7B, especially for memorization/recall. github.com/facebookrese...

visa and mastercard deciding what games are allowed to be sold is an apocalyptic level threat to the games industry. phone calls are the highest impact way to help.

this is an all hands on deck situation. the enemy got results with 1000 phone calls.

we can beat them

phone numbers and a script:

If you never thought it was possible to decrease a positive number by 1000%, think again!

The Commission is pushing for a massive increase in funding for Horizon Europe 2028-2034. If I understand correctly, Pillar I, which includes @erc.europa.eu and Marie Sklodowska Curie fellowships, would go from the current €25 billion to €44 billion. Let's hope it happens!

Update on hidden prompts in papers targeting LLM reviews: ICML 2025 PCs react.

icml.cc/Conferences/...

Historians Politely Remind Nation To Check What's Happened In Past Before Making Any Big Decisions WASHINGTON—With the United States facing a daunting array of problems at home and abroad, leading historians courteously reminded the nation Thursday that when making tough choices, it never hurts to ...

Historians Politely Remind Nation To Check What's Happened In Past Before Making Any Big Decisions — theonion.com/historians-p...

One like, one opinion about how to solve climate change.

A model trained with a frequentist objective using i.i.d. observational train and validation data is almost surely converging to the non-causal model of the underlying data generating process. That is because using short cuts/spurious correlations leads to a lower loss than not, i.e. you are actually specifically incentivizing your model not to be causal. "Emergence" believers' core fallacy is the assumption the causal model would have the lowest error, and therefore better optimization would eventually converge to the causal model. The opposite is true - the best scoring model (per frequentist objective) is actually the one that is best at exploiting spurious correlations.

Beautifully succinct statement of the core problem with the blind application of statistical learning in AI!!

A new study led by @timothysit.bsky.social reveals that different layers of mouse V1 integrate visual and non-visual signals differently.

Activity is dominated by vision (or spontaneous fluctuations) in L2/3 and by movement in L5. This leads to different geometries.
www.biorxiv.org/content/10.1...

JOB ALERT: Computational Cognitive Neuroscience Postdoc position in Osaka, Japan! Possible start in October 2025 (contact me ASAP), or from April 2026. PLEASE REPOST! #postdocjobs #neuroskyence #neuroscience #psychscisky #compneurosky #neurojobs 1/

I spent a while with Grok 4 trying to process what is truly leading performance, unlocking some new things AI hasn't done before, but with it a chronic (and increasing) set of system failures that make me question where the xAI project is heading.

buff.ly/cwHYbqn

Mechanical Watch – Bartosz Ciechanowski Interactive article explaining how a mechanical watch works.

Was chatting to one of my most brilliant Physics students yesterday and he had no idea how mechanical watches work and was fascinated by the fact that they make use of such basic Physics principles and the history of how watchmakers have overcome various limitations on accuracy. Shared this with him

👨‍💻 Open PI position in our institute 👨‍💻 !! If you are an expert in Computational Neuroscience and want to start your lab in Bordeaux, contact us !
www.fens.org/careers/job-...
@neuromagendie.bsky.social
@neurobordeaux.bsky.social

Impartiality in the media can be challenging at times, but it is certainly NOT achieved by simply both-siding every argument!

Especially when one side is a distinguished expert in their field, and the other side is an opinionated nincompoop.

#bothsideism