Francesco Nigro retweeted
AI may well be able to write massive software projects “from scratch”, but that almost certainly isn’t the fastest / highest performance / highest security / most token efficient way to do things. Good building blocks matter.
In addition to RL’ing the agents on final task loss, we should try to get them to regularize the building blocks as well. What would the standard library look like that makes the agents most effective?
It is likely at the limit that it turns out to be some alien organization of code, but given the existing human-bias in training, I bet that in the near term it would be lovely code that old programmers would appreciate.
I suspect the same could be said for mathematics, but that is out of my lane.
Francesco Nigro retweeted
Frameworks matter
AI may well be able to write massive software projects “from scratch”, but that almost certainly isn’t the fastest / highest performance / highest security / most token efficient way to do things. Good building blocks matter.
In addition to RL’ing the agents on final task loss, we should try to get them to regularize the building blocks as well. What would the standard library look like that makes the agents most effective?
It is likely at the limit that it turns out to be some alien organization of code, but given the existing human-bias in training, I bet that in the near term it would be lovely code that old programmers would appreciate.
I suspect the same could be said for mathematics, but that is out of my lane.
Francesco Nigro retweeted
Released fastconstmap 0.11.0 with free-threaded Python support (3.13t / 3.14t).
It is an immutable map from strings to 64-bit integers using binary fuse filters. It uses about 9 bytes per key (18 if you want missing-key detection), one xxhash plus three array reads per lookup, and a batched API for even more speed. It is interoperable with the Go and Rust versions of the same structure.
It will use massively less memory than the alternatives.
Because the map is immutable after construction, lookups need no locking. The extension now declares it does not need the GIL, so on a free-threaded interpreter several threads can look up in parallel.
Link: pypi.org/project/fastconstma…
New failure mode of LLMs: "Sorry, I trusted the doc without reading the code and derived from it the wrong behaviour. My fault: I have a rule to produce 0.6 docs per line of code"
WTF?!!!!
Francesco Nigro retweeted
You do need to learn a lot of new design principles that AI won't do for you - get a feel for how to avoid bottlenecks, reduce data dependencies, etc. There are also a lot of design patterns that are more important to SIMD (parallel prefix, use of compress and permute, ...
Francesco Nigro retweeted
@Steam is trying to be good with the AI label, but they are walking into a trap. The AI label is used to drive hate campaings against developers, just like a woke (SweetBabyInc) label would be used. Steam should reconsider. We don’t need more internet hate campaings.
Francesco Nigro retweeted
Hello you fine Internet folks,
Today's article is on Qualcomm's "Performance" Oryon AKA Oryon-M and how it compares to Arm's Neoverse N2 (the server version of the Cortex A710) and Intel's Skymont.
Hope y'all enjoy!
chipsandcheese.com/p/perform…
Francesco Nigro retweeted
gpt额度太多了 vibe了一个自用的玩意
多个agent memory不共享 但是我又不想把一些决策历史放入文档或或者git note也不想写进注释
搞了一个玩意 类似于idea的书签功能 可以记录某个文件 上若干行上的评论
类似于github的行级评论
支持 cli mcp desktop 甚至可以用native image分发
github.com/dreamlike-ocean/l…
Francesco Nigro retweeted
I've been reviewing a lot of code written by Opus 5.5, and it still seems to make a lot of bad software engineering decisions. The kind that will accumulate/calcify into unmaintainable software over time. IMO even the agents will struggle to maintain vibecoded software.
Francesco Nigro retweeted
Same thing with GPT-6.1-Sol. It overengineers things all the time. Some people like to pretend it doesn't matter, but I think it depends on where the code lies on the tech stack. It's easier to tolerate brittle high-level code than a brittle foundation.
Francesco Nigro retweeted
AI should never be allowed to decide the software architecture. That's programmer's job. You have to review and micro-manage to force AI to stick with the designed architecture. AI will diverge from it otherwise and will use that as justification to diverge even more later.
Francesco Nigro retweeted
No specific heat to anyone in particular, it's less than optimal that the first sightings of new chips come from these bimbo measurements (ooh, Geekbench) and not in contexts where you can, say, get proper throughput/latency/architectural measurements
wccftech.com/amd-zen-6-highe…
Francesco Nigro retweeted
As @P99CONF is over, you can watch my talk about multi-core concurrency p99conf.io/session/multi-cor…
Happy to chat about it after you've watched it
Francesco Nigro retweeted
Beta tester just found a synchronous network call in Petafab, causing random 250ms stalls (uncommon). Now that is fixed. 60 beta testers have been doing a great job.
Francesco Nigro retweeted
For io_uring nerds...
uringscope: eBPF tracer and analyzer for io_uring. Reconstructs per-request latency, io-wq punts, and batching from kernel tracepoints, then names the problem and suggests a fix.
github.com/rch0wdhury/urings…
Francesco Nigro retweeted
The Cost of Profiling in the HotSpot Virtual Machine
dl.acm.org/doi/10.1145/36790…
@java
