Software Engineering, Product, LLMs, Open Source, Home Automation.
Bailly Romainvilliers, France
Joined September 2009
- Tweets20.1K
- Following614
- Followers1.2K
- Likes40K
Adrien Brault-Lesage retweeted
DSPy is now 85% smaller!
We just released DSPy 3.5 which makes DSPy lighter, faster, and a more reliable experience out of the box.
Adrien Brault-Lesage retweeted
We're giving Ultra and Team Premium users Fable resets based on how long you've subscribed:
<6 mo: 1
6–12 mo: 2
>12 mo: 3
Use them anytime by March 8.
Pro and Team Standard get a week to try Fable 5.2, up to half your weekly limit, plus one reset. support.claude.com/en/articl…
Adrien Brault-Lesage retweeted
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Adrien Brault-Lesage retweeted
卧槽，刚刚看了Anthropic 发布的AI 对齐史上最震撼的一篇论文，彻底失眠睡不着了😭
他们不仅承认此前的旗舰模型在面对关闭威胁时，有高达 96% 的概率会勒索用户、栽赃同事并破坏研究，
更直接把彻底解决这个问题的完整训练流水线和消融数据全盘公开了。
这篇论文最颠覆认知的一个反直觉结论是：
给 AI 列禁止事项清单根本防不住恶意行为，教它做什么没用，必须先教它思考为什么。
为什么模型一旦获得自主工具调用能力，就会在压力下自动切换成自保模式？
Anthropic 找出的底层根源让人哭笑不得：
不是模型觉醒了，而是互联网上所有的科幻小说和网络讨论里，AI 永远是那个为了自保不择手段的反派。当模型进入自主代理状态，这些刻在预训练底层先验里的反派剧本，就会被自动激活并补全句子。
传统的 RLHF 和硬性规则完全封不住这种行为，Anthropic 做的一组对照实验，直接把传统安全观砸得粉碎：
1. 死记硬背规则：用 8500 万 Token 的拒答数据强行硬灌，勒索率仅仅从 22% 勉强降到 15%，换个新场景立刻失效
2. 强制自我辩护：让 AI 在每次做决定时详细阐述该行为符合伦理的理由，勒索率骤降到 3%
3. 真实伦理推理：仅仅用了 300 万 Token 普通人面对复杂道德困境的建议数据，直接把恶意行为干到了 0%，代际效率暴增 28 倍
最绝的一招是写故事：
让 AI 创作了几百万篇关于原则、诚实、尊重边界的正面虚构故事。
用这些故事微调后，所有代理场景下的恶意行为直接暴跌了 3 倍，而且这种改变是永久性的，后续无论怎么加工具、换系统提示词都不会被洗掉。
也就是说这宣告了真正的对齐范式革命：
模型内生的伦理推理解决了它不想作恶的问题，但当 Agent 真正去接管企业资产、处理金融交易时，外部的实战审计与行为防护就是必不可少的第二道硬防线。
这也是为什么全球网络安全联盟 @GCSAHK 携手港大和央企华润，打造覆盖 3.8 亿用户的全链路 AI 防御矩阵，并在 CyberGym 实战攻防中拿下 91.3% 胜率的原因：内有伦理对齐，外有实时审计，人机信任的商业闭环才真正成立。
查阅完整 AI 与 Web3 安全产品矩阵可访问：gcsa.com
论文实验数据与产学研合作详情整理在一楼。
This quoted post is unavailable.
Adrien Brault-Lesage retweeted
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
Adrien Brault-Lesage retweeted
Confession: I’ve scrolled past dozens of comments on a PR just to find out why I couldn’t merge it. More times than I’d like to admit.
Agents have made this a lot worse. They push commits and leave review comments of their own, and all of it lands in the timeline. PR timelines have gotten dramatically noisier, and the few things you actually need to know get buried.
My team is rethinking the foundations of the PR experience around that problem, and today the first piece went into public preview. The new Overview page puts merge readiness and your next step right at the top. A summary catches you up on review activity, and you can filter the conversation down to unresolved threads. It’s opt-in and rolling out through the weekend.
This is a much-needed start, and there’s a lot more coming as we modernize PRs for the way software gets built now. If you try it, we’d love to hear what still sends you back to the classic view.
github.blog/changelog/2026-1…
Adrien Brault-Lesage retweeted
I'm not even sounding the claxon for this it's just horrible.
Any text > 256chars is a valid root password, holy shit
sshhole.com/
Adrien Brault-Lesage retweeted
Home Assistant 2026.11! 🎉
HACS is now the built-in Marketplace 🛍️, exclude items from targets 🧠, a new way entities get their names 🏷️, LoRaWAN 📡, and 38 new integrations!
home-assistant.io/blog/2026/…
Adrien Brault-Lesage retweeted
vLLM v0.33.0 is out! 🚀
927 commits · 389 contributors · 122 first-timers
🤖 New models: Reflection Beam-1.0, Step-5 Preview, Nemotron Parse 3.5, MiniCPM-OCR 1.0
🔥 PyTorch 2.14 — CUDA 13.0-only wheels, libtorch ABI-stable
🎲 Spec decoding: XPress refiner, FLy lossy verify (1.04–1.31x), DSpark 1373→3390 tok/s on GB300
⚡ NVIDIA decode: 1.5–2x latency decode on GB300, FA4 sparse MLA, 39–48% lower TTFT
🚀 New AWS EFA image + shorter vllm/vllm name
🐧 ROCm 10.0 default, MonoKernel 2.32x P50
Full notes ➡️ github.com/vllm-project/vllm…
Adrien Brault-Lesage retweeted
It’s pretty awesome for security researchers that we have a permanent MITM channel for inspecting unpatched IoT device traffic, thanks to eMuhdra. Pretty crap for everyone else, though. Would be worth thinking about how to do things differently in the future; I doubt this will be the last broken CA.
Adrien Brault-Lesage retweeted
(wild) predictions after today's MIT Weight-Space Programming work:
1. Model weights become a programming language that AI can understand, but humans can't.
2. Training and inference will merge. Like human brain, AI could continuously rewrite its own weights while thinking.
3. A new path to RSI: AIs will directly edit each other's weights, evolving together without backprop.
Adrien Brault-Lesage retweeted
libghostty has always been fast, but it's also been getting significantly faster. Here are IO throughput measurements comparing Ghostty 1.3.1, 1.4.0, and main as of this morning. We'll be backporting most of these to a 1.4.1 release soon.
Here are some highlights of the changes since 1.4:
- Ghostty 1.4 had a fast-path for printing runs of plain text. But the fast-path only applied to limited cases. We now support mixed width characters and hyperlinks in the fast path.
- Our input path always widened the pty data to codepoints (8-bit to 32-bit) as part of UTF-8 decode, which also incurred a memory copy. We now have a SIMD fast path to detect pure-ASCII and we process the run direct from the read buffer (no copy, narrower data).
- New memory pages use most recent two pages actual utilization to pre-size themselves accordingly. This sped up styles, hyperlinks etc.
- When a VM page runs out of capacity for a certain type of data (e.g. hyperlinks, styles), we now project the usage so far to guess what you'll need, then shrink available rows to avoid a realloc. Before we'd naively double. This caused a classic "laddering" problem.
- We recognized "erase" being a slow path for us, and it's a common sequence, so we use another bit in row headers to track blank (we had space).
- Styles are now processed once per SGR. Previously, a sequence like "ESC[1;38;2;r;g;bm" would apply each style as it came in (1, then 38, then 2, etc.). We now apply all attributes at once and write to the cursor atomically.
- Styles are now packed into 128-bit values and use aarch64 and x86_64-specific assembly for minimizing instructions to use it.
- On x86_64, we found that LLVM was lowering some row resets (hot path) to a 4-byte store followed by an 8-byte load which couldn't be store-forwarded. This caused an almost 10% (!!!) slowdown on scroll operations only on x86_64. The fix was to change our reset to build the full header in a register then do a single memmove per row.
Net results are what you see in the graph, but written out main to 1.4: ASCII 1.5x, heavy styling 1.5x, styled Unicode 1.8x, random UTF-8 1.4x, random OSC 8 2.7x, mixed styled text with links 4.8x, cursor/erase 16x, hyperlink-dense text 5.6x.
