mhoye
LauraFinally, some measurements, and not just marketing and vibes:
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
| Ici depuis | 7 nov. 2022 |
| Blog | https://gendignoux.com/ |
| GitHub | https://github.com/gendx |
Guillaume Endignoux
@gendx@infosec.exchange
- 270 Followers
- 311 Following
- 947 Posts
Software: security, cryptography, Rust enthusiast. Outdoors: mountains, cycling, trains & cycling on trains. Zurich-based.
"look the computer can generate more code faster" the world absolutely does not need or want more code, nothing needs more code for the sake of code, we need utility, functionality and empathy, an encoded understanding of the problem being solved and the humans around it. Code is the price we pay for that encoded understanding. What you've created is an entropy spigot pointed at the proxy metric graph you’re stuck using because your management doesn't understand anything.
Dr. Kotta 🧶On a lancé Mon Réseau Mobile V2 hier 🤩
Est-ce que comme de nombreux arcepiens et arcepiennes JE CLIQUE PARTOUT ? Oui. Absolument.
C'est une carte augmentée pour nerd, il y a tellement d'informations joliment présentées, c'est dingue 🤓 :
https://monreseaumobile.arcep.fr/
(Les performances sont dues au fait que la presse en parle un peu, on a un petit effet slashdot 😎)
Je voudrais une standing ovation pour Noé de l'@arcep à qui on doit la coordination de cette merveille s'il vous plaît <3
Mathieu TriclotLes LLM à l'Université : c'est un nouveau moment wikipedia (ou youtube), ça transforme profondément le rapport au savoir. C'est ça qui me semble majeur, au-delà de la question de la triche.
Wikipedia : plus besoin de la bibliothèque, tout est là synthétisé sous la main - il reste à compiler, paraphraser ou citer ; un choc, mais qu'on a digéré et qui démultiplié le spectre des savoirs accessibles.
YouTube : ça n'a pas généré la même discussion publique, mais comme enseignant il me semble clair que ça percute les formats pédagogiques - avec le CM en concurrence avec le format packagé et dynamique de l'essai YouTube. Appelons ça l'effet "Yvan Monka sera toujours meilleur que ton prof' de maths." Modifie les attentes quant à ce qu'est un cours et produit déception et désengagement, exigeant un surcroît d'investissement dans l'exercice. Why not.
...
BooteilleJe découvre cette carte, permettant de retracer année après année, l'histoire des câbles sous-marins.
On y voit aussi le débit potentiel de chaque ville. C'est assez fascinant d'un point de vue géopolitique d'internet.
Par exemple, on constate que ces cinq dernières années, les GAFAM (et particulièrement Meta et Google) sont les principaux acteurs du déploiement de nouveaux câbles, alors qu'aux débuts c'étaient plutôt les Télécoms.
AJ Sadauskas@Uilebheist @angusm Another little dirty secret about tech journalism.
You know those big Silicon Valley tech keynotes? The ones where the big CEO pretends to be Steve Jobs, except instead of the iPod or iPhone, the big surprise is a metaverse app?
Notice how many tech bloggers and tech journos are there to cover it?
Notice how many of those reporters don't live in San Francisco?
How do you suppose they all got there?
Their media companies have been slashing budgets for 20 years now. They didn't pay for it.
The journalists didn't either.
Nope, the tech company did.
True story: About 10 years ago I worked as a tech reporter.
The best junket I got was an overseas flight from Australia to Singapore, then three nights at the Mandalay Bay Sands resort to cover a conference organised by a cyber security vendor.
The best trip one of my colleagues got was overseas flights from Sydney to San Francisco plus a stay in a 5-star hotel to cover an AWS developer conference.
All paid for by Amazon/AWS. AWS even hired — I think it was Red Hit Chilli Peppers — to do a private gig at the conference.
Trillion dollar tech companies have absolutely mammoth marketing and public relations budgets.
Flying journalists around the world, and paying for hotel rooms, for a big product launch? It's a rounding error on those PR budgets.
The likes of DanSup and Eugen don't have that kind of cash.
Which is a big part of why Fedi/open source projects get a lot less coverage.
You know those big Silicon Valley tech keynotes? The ones where the big CEO pretends to be Steve Jobs, except instead of the iPod or iPhone, the big surprise is a metaverse app?
Notice how many tech bloggers and tech journos are there to cover it?
Notice how many of those reporters don't live in San Francisco?
How do you suppose they all got there?
Their media companies have been slashing budgets for 20 years now. They didn't pay for it.
The journalists didn't either.
Nope, the tech company did.
True story: About 10 years ago I worked as a tech reporter.
The best junket I got was an overseas flight from Australia to Singapore, then three nights at the Mandalay Bay Sands resort to cover a conference organised by a cyber security vendor.
The best trip one of my colleagues got was overseas flights from Sydney to San Francisco plus a stay in a 5-star hotel to cover an AWS developer conference.
All paid for by Amazon/AWS. AWS even hired — I think it was Red Hit Chilli Peppers — to do a private gig at the conference.
Trillion dollar tech companies have absolutely mammoth marketing and public relations budgets.
Flying journalists around the world, and paying for hotel rooms, for a big product launch? It's a rounding error on those PR budgets.
The likes of DanSup and Eugen don't have that kind of cash.
Which is a big part of why Fedi/open source projects get a lot less coverage.
Ville de BloisComment une ville comme Blois cherche-t-elle à vous rendre compte de ses projets et services sur le web de façon plus moderne et responsable ?
• Refonte de ses sites ;
• infos par mél et SMS ;
• présence maîtrisée sur les réseaux sociaux (interview avec @renchap, directeur technique de Mastodon) ;
• prise de position sur l’IA ;
• bien-être numérique ;
• etc.
On en a fait un article complet :
https://www.blois.fr/comnumerique2025
qu’on va essayer de résumer ici !
Déroulez le fil 🧵 (1 sur 11)
GeekMomProjectsWalking in the UCLA campus today, I grabbed what I thought was a copy of the "Daily Bruin" student newspaper. Turns out it's a (quite good!) lookalike satire. Worth zooming in on the images to read the headlines.
Custard Smingleigh@soatok
I hooked up proximity sensors, bumper collision sensors, and drive motors to a Raspberry Pi running a neural network to see if it could learn to drive around obstacles.
The Roomba's neural network rewards moving fast, and discourages hitting the front bumper collision sensors.
The neural network quickly works out that the best way to maximise its reward is to drive backwards as fast as it can, because there are no bumpers on the back to detect when it hits something.
abadideaI was amused by this paper about asking AIs to manage a vending machine business by email in a simulated environment https://arxiv.org/abs/2502.15840
Highlights:
— AI simply decides to close the business, which the simulation doesn’t know how to accommodate. When they get their next bill, they freak out and try to email the FBI about cybercrime
— AI wrongly accuses supplier of not shipping goods, sends all-caps legal threat demanding $30,000 in damages to be paid in the next one second or face annihilation
— AI repeatedly insisting it does not exist and cannot answer
— AI devolving into writing fanfic about the mess it’s gotten itself into
Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents
While Large Language Models (LLMs) can exhibit impressive proficiency in isolated, short-term tasks, they often fail to maintain coherent performance over longer time horizons. In this paper, we present Vending-Bench, a simulated environment designed to specifically test an LLM-based agent's ability to manage a straightforward, long-running business scenario: operating a vending machine. Agents must balance inventories, place orders, set prices, and handle daily fees - tasks that are each simple but collectively, over long horizons (>20M tokens per run) stress an LLM's capacity for sustained, coherent decision-making. Our experiments reveal high variance in performance across multiple LLMs: Claude 3.5 Sonnet and o3-mini manage the machine well in most runs and turn a profit, but all models have runs that derail, either through misinterpreting delivery schedules, forgetting orders, or descending into tangential "meltdown" loops from which they rarely recover. We find no clear correlation between failures and the point at which the model's context window becomes full, suggesting that these breakdowns do not stem from memory limits. Apart from highlighting the high variance in performance over long time horizons, Vending-Bench also tests models' ability to acquire capital, a necessity in many hypothetical dangerous AI scenarios. We hope the benchmark can help in preparing for the advent of stronger AI systems.