Les tweets "j'ai essayé l'IA et ça m'a fait de la merde, c'est bien la preuve" me font penser aux tweets "on parle de réchauffement climatique, mais tu as vu la température aujourd'hui ? Rien à voir !". Pas plus lucide, pas plus honnête.
There’s no excuse for new software not to be logical, not to be deterministic.
Investing in owning your testing, to be DST-first, and to accelerate the local development velocity of your team, is probably the best technical decision you can make.
I refuse some of the most important real-world systems in aviation cannot handle 200 QPS
It’s almost an inconsequential amount of traffic for any modern system
While I understand aviation is not considered “modern” by tech industry standards - come on…
Revolut shared the postmortem of their global cards outage of 5 hours. Some interesting (alarming?) notes:
1. They had no monitoring on database nodes going down (!!)
2. Across 3 regions, they ran 5 nodes, with #1. So a single node going down meant no zonal resillience (cont'd)
We've used TLA+, Jepsen, and chaos testing in TiDB, and I still worry about what we failed to model. 😂
AI makes writing formal specs easier. But who verifies that we're specifying the right thing?
Now we can formally prove the wrong thing 10x faster.
Welcome to vibe verification. 😂
Oh boy, this sounds even more like a failure with a feedback mechanism. Time to find this feedback loop and break it. Most (but not all) loops need an overload on some physical or logical resource somewhere. @Microsoft I am on sabbatical to work on all kinds of metastable stuff😉
I strongly believe (hope?) humans should design and maintain software (likely in natural language). As this transition happens, we must also increase the
level of abstraction used for verification (aka tests)
Transcript Tests can help: https://t.co/BPTk1nYpFC
@gwenshap Spending the effort on the right abstractions and layering, then letting DST battle-test the code, works surprisingly well. And when something breaks, you get a seed to replay the exact failure deterministically as a feedback loop
En bouclant cet article (https://t.co/f7nY016woD ), il y a quand même un truc qui me rend triste. Pas pour Jensen Huang lui-même, mais pour Nvidia en tant que geek
Toute la magie (et la folie) de Nvidia tenait dans une anomalie : une boîte de 4 000 milliards pilotée comme une start-up en mode survie, sans numéro deux, sans comités, où un seul cerveau tranchait les paris d’architecture à dix ans
Aucune PME ne survivrait à un audit avec une telle gouvernance, mais c’est précisément ce qui a donné CUDA et la révolution de l’IA
Pour survivre à son fondateur, Nvidia va maintenant devoir se normaliser, créer des étages hiérarchiques et ralentir. Et quand les purs gestionnaires remplacent les ingénieurs obsessionnels, l’histoire de la tech montre qu’on y perd toujours quelque chose...
I started Barrel a few years ago as an alternative to CouchDB. Put it aside for a while, then came back to it with a broader direction.
A database for AI agents, with documents, files and vectors managed together.
Built in Erlang, open source.
https://t.co/asQ4HN3dGy
Agentic coding is a form of machine learning. Generated code is best treated as a blackbox artifact whose behavior and generalization should be managed via empirical evaluation, like with any ML model.
@meln1k My best strategy is to ask for buggify nodes each times a rare seeds is found, to be sure the situation is happening often
https://t.co/hgTpZRFqFk