benoît chesneau

My GitHub Sponsors profile is live! You can sponsor me to support my open source work github.com/sponsors/benoi…

you take the time to check logs, check some AI analysis (yes we are in 2026), then lost your time to redact, and re-edit to clarify. But the other keep answering using AI or suddenly became verbose in english with a true picture in mind and key points to show.

A few more words about Barrel on LinkedIn: linkedin.com/posts/benoitc_…

Introducing Barrel 1.0: a local-first database for AI applications and agents. First released seven years ago, Barrel is now back after a long hiatus. Built on Erlang/OTP, fully open source, and licensed under Apache 2.0. barrel-db.eu #Erlang #BEAM #AI #LocalFirst

not that bad - docdb: read 30.8K ops/s, update 41K ops/s; indexed+limited queries in tens of µs (prefix_limit 17K ops/s), unindexed multi-condition scans in ms; subscription 12.6K ops/s. - vectordb (HNSW): insert 13.2K ops/s, search sub-millisecond at p95 across k1/k10/k50.

@gargu_ also don't get me wrong, we still need companies able to do the data curation and it will have a cost. But besides that having bgp netflow generalized with a good protocol that allows per AS to distribute the protection to the edge should be a standard offer at some point imo

@gargu_ it only cost this because operators don't coordinate. anti ddos sinks just remove the principal of internet which was about decentralizing services. current infrastructure makes internet very fragile

ddos protection should be something offered by each operator. We shouldn't rely on giant sink like F5 and others to do it.

@bluetouff github.com/barrel-platfor… sinon

Fuyez ce truc, montez vos propres serveurs MCP, ne créez pas un point de centralisation de la monétisation des contenus ou de la distribution de contenus agentiques. Mon serveur MCP est libre et vous pouvez l'utiliser comme bon vous semble, même pour des projets commerciaux : github.com/bluetouff/l0g/…

$ 94 000 is far from 0.

@jedisct1 you would imagine they would have plane a migration step instead of letting people on their own...

macOS 28 will no longer support encrypted Mac OS Extended (HFS+) volumes support.apple.com/en-us/125615

📣 The moment is here. 📣 TypeScript 7 is officially released! 7️⃣ devblogs.microsoft.com/typescript/ann…

Announcing Robostral Navigate, our first model for embodied navigation: an 8B robotics navigation model that guides robots to autonomously perform tasks specified with natural language. Single RGB camera. State-of-the-art on R2R-CE.

Imagine if the money spent on GPUs, low-precision inference, and AI influencers went into founding AI labs across Europe instead.

just like software moved beyond shipping binaries everywhere...

"LLMs are just files that can be copied." Sure they are. That's like saying you want to live with an immutable snapshot you'll have to keep forever. We need something new. We need cheap hardware. ASAP.

@undefinedKi @N01ennn uou will hear it's not unreel

@N01ennn pairing two units to run 405B models on a desk still sounds unreal

TWO NVIDIA DGX SPARKS CONNECTED DIRECTLY OVER A 200 GIGABIT LINK POOL 256GB OF UNIFIED MEMORY FOR $7,998 TOTAL, ENOUGH FOR EVERY OPEN WEIGHT MODEL UNDER 300B PARAMETERS WITHOUT QUANTIZATION 00:22 he holds the board up to the camera, "you get NVLink for unified system, which gives you shared access to that 128 gig of memory, which is really important" each DGX Spark ships in a 1.13 liter chassis with the GB10 Superchip, a 20-core ARM CPU, and 128GB of coherent memory. two 200 gigabit ports sit on the board next to a NIC miniaturized down to postcard size the interesting part is the pairing. plug two units into each other over the 200 Gb link and the OS exposes 256GB of unified memory across the pair. run vLLM, shard the model, no external switch required for local AI this is the first time frontier-class inference lives on a desk instead of a rack. one DGX Spark runs 200B parameter models. two paired run up to 405B. add a 200 Gb switch and you can scale to trays of them feeding from data center storage over NVMe over Fabrics the one design flaw is the 2242 SSD, not rated for AI workloads. NVMe over Fabrics to external storage is the intended path the article ranked the $4,199 Mac Studio as Device 4. two DGX Sparks at $7,998 sit above that as the first practical multi-node local AI cluster available to a single buyer bookmark this and read the article below

ces cafes qui ont un minimum pour les cbs...

@0xSero what's the breakeven of it?

I'm buying a second DGX Spark, my goal is very simple. In daytime I will have DS4 Flash for chat running across 2x Sparks and GLM-5.2-REAP running on 4x 6000s I am setting up a goal builder, which recommend me loops to run overnight. At night I will have GLM-5.2 (not REAP) running with PP = 2 with stage 1 = TP4 with 6000s and stage 2 being TP2 with the DGX Sparks. To make this work I need to get really creating, but I think I can get 25-40 tok/s decode for the Nvidia NVFP4 At night I don't care much about tok/s as long as it'll be large enough to make progress over 8 hours.

raises an hand if you start hearing people saying "so I will have the *full picture*" . claudification in progress.