Mise à jour de profil! 🕊️ J'ai cofondé avec @antoine_peze Synopsis ➡️synopsis-studio.com Synopsis porte la plume et la voix de vos clients partout dans votre entreprise. Des enseignements actionnables pour que vous ravissiez vos clients et construisez une marque forte. 🪶

Two decisions this spring should be read as one warning. One came from Beijing, one from Washington. Both tell Europe the same thing about its place in the technologies that now decide strategic power.

Oh, everyone wants Europe to do well on AI, easy to say, but how many of you are using Mistral? Words are cheap, a slightly worse model that you have to fine-tune, where the batch mode jobs fail 20% of the time without refund, and suddenly to yank out on us? Pathetic.

So… did everyone uninstall their IDE or what ? Have we forgotten how to explore a codebase, moving around taking notes, and plan refactorings ?

Le réflexe de l'appel téléphonique est symptomatique d'une époque qui avait un travail beaucoup moins intense, beaucoup moins continu. Autrement dit: beaucoup plus de temps libre à allouer dans sa journée.

Been thinking about what an "agent-native cloud" actually needs to look like. Mentioned this, and @Vercel's CEO replied that it'll be them. Cool! Here's the spec they (or @Cloudflare, or some startup not yet invented) actually have to hit. It won't be @awscloud. Thread...

Retarded benchmark. What is its purpose ? This is so backwards regarding the role of software and how it’s built. I hope this bench will flop

So DeepSeek-V4: finally took me the week. Overall the paper is attempting many things at once, not easy to disentangle as it's all surprisingly connected. It's first a serious attempt at briding the gap between close and open LLM architecture. It is generally rumored that Opus and [largest model bundled in GPT-5] belong to an entirely different category of models: very large, very sparse mixture of experts, able to holding an unprecendently wide search space while still being servable. Simply put current hardware cannot hold a model on one node, so you have to play with the interconnect and various level of quantization, for different layers, at different stage of training. An important focus of DsV4 is on communication latency, showing it can be hidden through effective management of interconnect (roughly you slide communication time inside computation side). Overall, you cannot simply enter this game without the capability to rewrite kernels from scratch and the model report relentlessly come back to this. Because this is the frontier game. It's then a radical, but very successful attempt at making long context simultaneously more efficient and more affordable. Long context is literally a "context" problems: what exactly is worth attending? An obvious fix is to prioritize the most recent tokens. This might be sufficient for basic search but not for the new demands of agentic pipelines that require accurate recall of distant yet strategic content. V4 clever approach is to rely on two different axis of memorization by allocating layers to two different attention compression schemes. Like the name suggest, Heavily Compressed Attention is the brute force method collapsing each sequence of 128 tokens to a unique entry and take care of the fuzzy yet global context. Compressed Sparsed Attention rely on a "lighting indexer" to bring the relevant local blocks for query, even when they can be thousands of tokens away. Everything here is optimized for end inference: there is very large head_dim (512) which is costlier for training but allows for even more compressed kv cache which is your actual bottleneck at inference time, especially in prefill mode. End result is very classical DeepSeek play, introducing a new radical disruption of inference economics after DSA. I predict hybrid CSA/HCA (or similar counterparts) will be essentially part of the mainstream arch by the end of this year. Now we come to the more ambitious but also more unfinished part: an attempt at redefining model architecture and the learning signal. Most preeminent part is mHC and hybrid CSA/HCA, but it's actually a long list of less documented innovations: swapping softmax for sqrt(softplus) or using an hybrid two-stage scheme with non-standard values for Muon. Yet the interconnection all of these new components is still unknown and likely to account for the significant training unstabilities: typically "mHC involves a matrix multiplication with an output dimension of only 24" which introduces non-determinism. Even one the best AI labs in the world will run here into ablation combinatorial explosion, so the association of all these choices is likely non-tractable and would require a more consistent theory — which the conclusion gestures at, but does not solve ("In future iterations, we will carry out more comprehensive and principled investigations to distill the architecture down to its most essential designs"). The more limited experiments in post-training are maybe more promising. Significantly, the one lab that popularized the standard RL reasoning recipe is rethinking the recipe. For now it's a two stage design (RL on specialized model, then on-policy distillation): ever since Self-Principled Critique Tuning DeepSeek has been concerned with expanding the reasoning training signal beyond final sparse reward. I'm not sure this is final say: in this domain everything is a bit in flux and you could even argue the type of verified pipeline we designed for SYNTH is a form of extreme offline RL-like training. There is an even longer term plan (here >3-5 years), which is about redefining hardware. For now it's a way of transforming a constraint into an opportunity: as the leading Chinese labs, DeepSeek was very incentivized to make training work on Ascend and contribute to the national effort for chips autonomy. Very unusually, the report includes a lengthy wishlist for future hardware to come in the report itself. As several experts noted, many of these recommendations don't really hold up for Nvidia but make perfect sense for a newcomer in the GPU hardware business. DeepSeek seem to be anticipating a world where labs have to secure a close hardware partner to retroactively fit the chips to the particular demand of model design or inference. Now there is what DeepSeek did not do yet. The paper hardly mention anything about synthetic pipelines, rephrasing, simulated environment. Training data size (32T tokens) likely involve some significant part of generated data, as this is more quality tokens than the web and other digitized sources could held — so maybe similar synthetic proportions as Trinity (roughly half) or Kimi. Still, it's pretty clear that all their attention was focused on the infra, architecture and scaling side, leaving a proper extensive retraining for later. This is likely not that dissimilar to how Anthropic or OpenAI proceeded: the fact we're still in the middle of the same model series even though significant parts of the model have changed (the tokennizer with Opus 4.7) suggests that a model lifecycle involves multiple rounds of training potentially as large as a pretraining a few years ago. The fact DeepSeek took on multiple Moonshot innovation (and Moonshot in turn has been hugely reliant on DeepSeek) suggest we might also have an ecosystem dynamic here. Maybe DeepSeek can exclusively focus on hard infrastructure problems and expect some of the axis of development to be sorted out later.

I can more or less agree with your assessment. However there are obvious technical constraints and moreover programmation constraints. The deal was thought of this way: Germany, which has high competence on heavy tanks, would handle MGCS, with minimal French involvement. France, which has high competence on fighter planes, would handle FCAS with minimal German involvement. 7 years later, here's how it turned out: - Germany demanded dominance on the MGCS design, and we accepted that they would get the lion's share of the program and base it off Leopard 2 and not Leclerc. In exchange for essentially giving up a French next gen tank, we asked for the ASCALON gun to be integrated, leaving the Germans with the entire frame, engine, design, and really everything else - Germany refused to integrate ASCALON and wanted full control of everything - Germany kept the project "going" very slowly and weakly for years and years - Germany ended up "having a baby in our back" and decided to soft-drop MGCS and create a German-only Leopard 3, leaving us holding the bag One can make a case that this is just how it turned out and that France should just buy Leopard 3s because France wasn't all that interested in making a heavy tank in the first place. But even then, we shook hands on a project, got nothing, waited for years, then were left to hang. Anyone would find that disrespectful. 7 years later for FCAS though, things are a million times worse: - Germany also made us wait 7 years - Germany tried to pull EVERY bit of construction in Germany despite France having the skills - Germany had "Airbus Spain" (nothing more than a subsidiary of Airbus Germany) incorporated to make it a "3-way partnership" - Germany and "Spain" aka Germany 2 then said "we represent 66% of the partnership so our majority decision should be respected" and every single decision was about demanding French technologies transfered to them and building plans being made in Germany and "Spain" - It was obvious to anyone with eyes that since Spain doesn't have the proper toolchain, the Spanish plans would all fail, at which point "Airbus Spain" would concede to send everything they have to Germany - Once Germany would have had nearly all production lines in Germany, they would have said "since we have almost everything working here, the problem is France needlessly holding on to their inefficient production lines" - Germany tried to argue every single last thing even when it was in the contract, like naval version of the plane which is a mandatory part for AC capability And key thing, this is EXACTLY the same as to what happened with Eurofighter. Germany "signed" with the rest of Europe then demanded to remove the engine to shove a bigger, less efficient and pointless German engine so they could make money. They tried for years to change the standard radar for one of theirs. They also, and that's the key problem here, held back and stomped their feet on every single last upgrade the plane needed and negotiated every single piece. This is why Eurofighter is, frankly, a shit. Why it got its radar 10 years late. Why it is only good on paper but in effect France alone has sold about 300 Rafales abroad while the entire consortium of England/Germany/Italy/Spain has barely sold half of that. Germany is perfectly content with bombing projects it is privy to, it has no qualms demanding other people's technologies, it has no problems slowing everything down to get as much as it can. In one word, they behave like arabs trying to negotiate everything and care very little about doing a good project. This pattern repeated itself on the Boxer (overweight, unwieldy), the Eurofighter, or that silly giant turret thing the Skyranger. Germans like to build big and expensive rather than efficient. This is not useful in war. Even the Ukrainians who tested Leopard 2s said "nice and all but we can't repair them and those older Leopards from the 70s actually are good". I agree that FCAS was probably a stupid project in the first place and that France, which has genuine ambitions for the plane, wants Dassault to run the project because when Dassault was given full control of the NEUROn, despite being an international project, it came out on time, was an excellent drone, and completely did everything at cost and with competence. Same happened with the SCORPION trio, Griffon-Serval-Jaguar, one boss, one goal, great success. One boss should run things. The Germans hate having a boss because they'd rather re-re-re-re-re-renegotiate everything, and the dirty tactics to "renegotiate the main contract" then blame us for "not accepting their reasonable demands" is just insulting. These people are not honest, and all their grubby attempts to take our technology are extra insulting.

Répétez avec moi: c'est le système (entreprise, équipe, etc.) qui est productif, pas les individus. Oui, la métrique "productivité" est individualisée, mais *n'est pas spécifique à un·e travailleur·euse* !

This is 100% true. IMO: Trump's historical function imo is that exposes any system he's in. As a real estate investor, he exposed the greed of the banks that lent money to him by playing them for fools. As a TV show host he exposed the cynical clawing ambition of people who work for him by showing how they would step on anyone for a job. As president he has exposed, as this tweet says, what America wants to let itself be, what cruelty we will allow as a country. He has also exposed the hypocrisy of Europe and Israel. Trump doesn't do this intentionally nor does he care about justice. He participates in the evil and by participating, he shows how ugly these systems are.

When the world does your French response for you.

This is what I've been saying since the summer. We are in the middle of software engineering's productivity crisis. LLMs inflate all previous productivity metrics (PRs, commits) without a correlation to value. This will be used to justify layoffs in Q1/Q2 of this year.

Dsl (merci aux amis pour les idées terribles)

Quick question, if you think it is impossible to grind them, because they weighted tons, how do you think they stacked them? Doesn't take a lot of energy too to put one block on top of another?

Je ne sais pas si cette retraitée est représentative de l'ensemble des seniors mais mon expérience ici et sur Facebook m'incite plutôt à croire qu'elle l'est. Par conséquent, je continue à penser qu'il est illusoire de croire qu'il est possible de convaincre les seniors de faire volontairement un effort qu'ils refuseront catégoriquement. En plus, nous n'avons pas le luxe financier d'attendre de faire toutes les autres optimisations possibles avant du type "réforme de l'Etat" ou "moins d'allocations pour les étrangers". Conclusion logique, nous perdons notre temps à tenter de convaincre des personnes qui ne changeront pas d'avis au lieu de passer du temps à réfléchir à ce qu'il faudrait faire. Je pense sincèrement qu'il faudra à un moment que le politique passe en force pour réduire d'autorité les pensions, probablement sous une pression externe, comme en Grèce à l'époque. Je vais être très cash, je me moque royalement que cela soit considéré comme antidémocratique et je rigole d'avance en pensant à des seniors qui brandiraient la menace du mouvement social. On ne m'a pas demandé mon avis pour mettre en œuvre un système financièrement insoutenable qui ponctionne 30 % du brut des actifs donc je ne vois pas pourquoi on devrait demander l'avis aux seniors pour leur baisser leurs pensions et attendre qu'ils daignent accepter de faire un effort. Aucun contrat social, et aucune promesse ne sont valables lorsque les paramètres de départ étaient viciés. Enfin pour que cela soit clair, je ne les tiens pas pour responsables de la situation actuelle mais je n'aurai aucune sympathie particulière à leur égard quand ils seront enfin touchés. Il faut arrêter de payer pour le passé alors qu'on devrait investir dans l'avenir et pour les plus jeunes.

1Password ➡️ @dashlane , obviously. It's not a joke, it's better than any other password management suite.

IL A RAISON

Many criticize the French, calling them arrogant or annoying, yet they are often the first to defend Europe. They sent a carrier when Greece was threatened by Turkey, troops to Romania after Russia attacked Ukraine, and now they deploy Rafales in Poland.