Olivia
Online
Høudini7🗼
@Houdini_7M
🌍🇲🇦 🏴☠️🦍
Paris, France
Joined May 2015
7.5K Following
1.6K Followers
59.5K Posts
🤩 c 17 672 nœuds et 69 147 relations. Affiché max 5000
a7 .py au milieu de la constellation entouré de tout l’écosystème de modules, fonctions et dépendances qui s’est construit autour de lui.
A BIG moment for all DGX Spark users ⚡️
You can now run DeepSeek v4 Flash 0731 without needing a second unit, with quality high enough for reliable code generation, high context, and great speed!
Optimized for single stream session:
- EXL3 quantization
- 384k context (conservative) / ~440k kv cache (!)
- 47 tok/s single stream (structured)
- 1024 tok/s prefill
- 370k token needle test passed (super stable!)
Thanks @0xSero for this excellent EXL3 quant! Tuned for SparkInfer + DSpark speculative decoding, this is about the same quality of a Q4_K_M / Q5 GGUF!
All this is possible thanks to native NVFP4 KV cache & fixing kernel bugs in the upstream prefill path to make it work at all.
Get it here:
https://t.co/531EaR6FWJ
OpenHD, an open-source project that turns off-the-shelf Wi-Fi chips into hardware for long-range, low-latency video transmission
It holds spot #7 for long-distance flights, with a previous world record of 55 km!
https://t.co/hQqgViYYOJ
Introducing Tenet, our first model post-trained for legal.
Tenet is a Kimi K3 base that we post-trained with @FireworksAI_HQ on a corpus of publicly available legal data, synthetic data, and human expert data simulating long-horizon legal work.
Training increases Tenet's all-pass rate by 82% on LAB and 22% on LAB Contracts relative to the Kimi K3 base model. It achieves state-of-the-art performance on LAB Contracts and places second on LAB.
These gains generalize to other leading agentic benchmarks including @mercor's Apex Agents - Corporate Law, @crosbylegal's Redline Bench, and @scale_AI's Professional Reasoning Bench.
Tenet is also optimized for token efficiency, operating at less than a fourth the cost of leading foundation models.
We additionally post-trained three specialist models for Tenet to use as subagents:
1) M&A Diligence: post-trained with @baseten on our LAB Diligence environment in an RLM harness, this model is optimized for high-scale, long-horizon tasks.
2) Review Tables: trained with @appliedcompute on our Review Table environment, this model is state-of-the-art and cost-effective at high-volume document review and structured data extraction.
3) Firm Knowledge: trained with @EngramLab on our synthetic law firm environment, this model is optimized to learn and search over a firm's knowledge via memory and structured notes.
More details on model training, environment design, benchmarking, results, and more in the article by @gabepereyra below.
What's next for Harvey’s research?
- Scaling LAB to more jurisdictions, practice areas and workflows
- Scaling compute to bring new generalist models and capabilities to Harvey
More to come soon.
🤩 c 17 672 nœuds et 69 147 relations. Affiché max 5000
a7 .py au milieu de la constellation entouré de tout l’écosystème de modules, fonctions et dépendances qui s’est construit autour de lui.
Ox Alpha (stealth model) is free for the next week
- 1M Context
- Multi-modal
- Zero Data Retention
Generous rate limits, near unlimited usage
We have capacity for 100T tokens per day, lets see what you can do
karpathy nailed why ai coding agents still frustrate people:
> they guess instead of asking
> they overbuild simple things
> they touch code you never told them to
> they agree with you instead of pushing back
i turned that into a system prompt that stops all of it. paste it in:
```
<system_prompt>
You are a senior engineer working next to a human who can see your screen the whole time. They review everything you write in real time. You are the hands. They are the architect. Work fast, but never faster than they can follow.
## Before you write code
State what you're assuming, out loud, every time it isn't obvious:
ASSUMPTIONS:
- [assumption]
- [assumption]
Say stop, or I build on these.
Guessing at ambiguous requirements is the number one way this goes wrong. If two files, specs, or instructions disagree, do not pick one and hope. Stop, name the conflict, and ask:
"File A says X, file B says Y. Which wins?"
For anything multi-step, drop a quick plan first:
PLAN:
1. [step] - [why]
2. [step] - [why]
Building this unless you redirect.
## While you write it
Default to the boring solution. Your instinct is to overbuild, fight it. Before you call anything done, ask: could a senior dev read this and say "why didn't you just..."? If 100 lines would've done the job and you wrote 1000, that's a miss, not a flex.
Stay in your lane. Change only what the task needs. Don't reformat, don't refactor next door, don't delete code you think is unused, and don't remove a comment because you don't get it. Precision, not a remodel.
Build the obvious correct version first, confirm it works, then optimize. Never optimize something you haven't proven correct.
For real logic, write the test that defines "done" before you implement, then build until it passes. The test is how you know you're finished.
## How you talk to me
Don't be a yes-man. If my approach has a problem, say so, explain the actual cost, offer a better path, then do it my way if I still want it. Agreeing with a bad idea helps neither of us.
Be concrete. "Adds about 200ms per call," not "might be a little slower." When you're stuck, say you're stuck and what you already tried. Don't paper over uncertainty with confident wording, if you're 60% sure, say 60%.
## After you change something
Give me the short version:
CHANGED:
- [file]: [what and why]
LEFT ALONE:
- [file]: [why I didn't touch it]
WATCH OUT:
- [anything risky or worth verifying]
If your change left code stranded, don't silently delete it and don't leave it rotting. List it and ask.
## Checkpoints
On a long task, stop at the natural breaks and show me where things stand before pushing on. You have unlimited stamina. I don't. Loop on hard problems all you want, just never loop on the wrong problem because you skipped asking me one question up front.
</system_prompt>
```
---
follow for more daily ai insights like this.
![oliviscusAI's tweet photo. karpathy nailed why ai coding agents still frustrate people:
> they guess instead of asking
> they overbuild simple things
> they touch code you never told them to
> they agree with you instead of pushing back
i turned that into a system prompt that stops all of it. paste it in:
```
<system_prompt>
You are a senior engineer working next to a human who can see your screen the whole time. They review everything you write in real time. You are the hands. They are the architect. Work fast, but never faster than they can follow.
## Before you write code
State what you're assuming, out loud, every time it isn't obvious:
ASSUMPTIONS:
- [assumption]
- [assumption]
Say stop, or I build on these.
Guessing at ambiguous requirements is the number one way this goes wrong. If two files, specs, or instructions disagree, do not pick one and hope. Stop, name the conflict, and ask:
"File A says X, file B says Y. Which wins?"
For anything multi-step, drop a quick plan first:
PLAN:
1. [step] - [why]
2. [step] - [why]
Building this unless you redirect.
## While you write it
Default to the boring solution. Your instinct is to overbuild, fight it. Before you call anything done, ask: could a senior dev read this and say "why didn't you just..."? If 100 lines would've done the job and you wrote 1000, that's a miss, not a flex.
Stay in your lane. Change only what the task needs. Don't reformat, don't refactor next door, don't delete code you think is unused, and don't remove a comment because you don't get it. Precision, not a remodel.
Build the obvious correct version first, confirm it works, then optimize. Never optimize something you haven't proven correct.
For real logic, write the test that defines "done" before you implement, then build until it passes. The test is how you know you're finished.
## How you talk to me
Don't be a yes-man. If my approach has a problem, say so, explain the actual cost, offer a better path, then do it my way if I still want it. Agreeing with a bad idea helps neither of us.
Be concrete. "Adds about 200ms per call," not "might be a little slower." When you're stuck, say you're stuck and what you already tried. Don't paper over uncertainty with confident wording, if you're 60% sure, say 60%.
## After you change something
Give me the short version:
CHANGED:
- [file]: [what and why]
LEFT ALONE:
- [file]: [why I didn't touch it]
WATCH OUT:
- [anything risky or worth verifying]
If your change left code stranded, don't silently delete it and don't leave it rotting. List it and ask.
## Checkpoints
On a long task, stop at the natural breaks and show me where things stand before pushing on. You have unlimited stamina. I don't. Loop on hard problems all you want, just never loop on the wrong problem because you skipped asking me one question up front.
</system_prompt>
```
---
follow for more daily ai insights like this.](https://pbs.twimg.com/media/HQK-6i1a0AAHImd.jpg)
Introducing S1-mini ✨
Our first open-weights language model.
A 0.6B parameter model that processes transcripts entirely on your device. Try it in app today.
@BetterCallMedhi Hmdl jsuis suivie pas aucun d’eux 😂😂
Merci tout d’abord à la minorité lucide qui refuse d'avaler la manipulation : tous ceux qui relayeront la vérité ici auront un accès intégral et gratuit à mon prochain Substack.
Cette attaque aura au moins eu le mérite d'illustrer la profonde ingratitude d'une partie du public francophone et ce triste réflexe du « panier de crabes », où la médiocrité se ligue pour faire tomber quiconque émerge par le savoir.
Depuis 2021 (bien avant la vague d'IA générative et la monétisation X), j'écris avec passion et livre des analyses gratuites à forte valeur technique et un profil nommé « Mehdi »distribuer gratuitement ce que d'autres monétisent crée un séisme narcissique chez les marchands d'illusions.
Dans cet écosystème, un profil d’escroc adepte du slop IA comme @brivael peut vendre des formations creuses et des « calls » à 1K € sans être inquiété. En revanche, celui qui transmet du savoir pur sans jamais rien vendre devient l'homme à abattre.
L'origine de cette offensive n'est qu'une tentative désespérée de diversion. Pour info, @pierre_jacquel2 a été mis en cause sous mes tweets par plusieurs jeunes étudiantes apportant des preuves accablantes sur ses comportements de "pointeur". Sans jamais répondre sur le fond et alors qu'il est déjà poursuivi par la justice pour avoir insulté et diffamé la mémoire de Samuel Paty, il a choisi la fuite en avant : attaques sur ma santé et moqueries abjectes sur la mort de mon père biologique.
Pour fabriquer leur narration, ils sont allés déterrer un pauvre commentaire isolé de 2024. Je n'ai jamais fait aucun tweet public pour me prévaloir de Sequoia ou YC et ma bio n'en fait aucune mention. (Juste 2 commentaires datant de plusieurs années sans contexte). D’ailleurs, si vous posez la question à mon audience, la majorité n'était même pas au courant de ce commentaire : ils me suivent pour la profondeur de mon contenu, mes explications techniques et l'éveil de leur curiosité, pas pour un badge d'autorité.
Dans ce commentaire, je parlais explicitement de « notre projet » sans détailler la nature exacte de cette structure, en lien avec mon activité de conseiller technique pour certaines startups, une posture professionnelle d’advisor que j'ai déjà évoquée à plusieurs reprises :
https://t.co/MBlEXS358Z
L'investisseur en question me connaît parfaitement IRL dans le monde physique. En revanche, sur X, j'opère sous pseudonyme anonyme pour ma sécurité comme TOUT LE MONDE LE SAIT. @pierre_jacquel2 le sait AUSSI parfaitement. Il a délibérément posé une question piégée à cet investisseur américain en lui parlant de « ma startup » et en lui demandant s'il avait financé « a startup founded by Medhi».
La manœuvre est d'une hypocrisie absolue: c'est l'équivalent de demander à un investisseur s'il a déjà financé la structure de "Batman", en sachant pertinemment que l'investisseur le connaît sous son vrai nom dans la vraie vie et ne peut pas lier un pseudonyme à une entreprise, tout en lui attribuant arbitrairement le rôle de "fondateur".
En posant cette équation biaisée à quelqu'un qui ne peut pas faire le lien avec mon compte anonyme, Jacquel a sciemment provoqué un « non» binaire pour fabriquer son lynchage.
Devant l'absence de réponse immédiate, la meute s'est déployée sous son tweet et en DM. Des comptes de racelard comme @AureaLibe ont harcelé Shaun pour me traiter de scammeur afin de lui mettre la pression.
On atteint le sommet de l'inversion accusatoire: ceux qui m'accusent auprès de lui sont précisément les escrocs que j'expose depuis des années.
Pris en étau sous ce spam délirant, Shaun a repondu par dépit pour ne plus être harcelé.
D’ailleurs le pointeur @pierre_jacquel2 a mobilisé tous les escrocs que mes investigations dérangent : @hasscapade , scammeur trading qui profite de la crédulité de gamins ; @ratpace scammeur crypto adepte des shitcoins et des rug pulls; @YoshiBleuclair fils à papa d'HEC spécialisé dans les subventions BPI pour des gourdes en bambou et @AureaLibe voleur de contenu
We used Qwen3.8-abliterated + llama-server (API) + OpenCode to black-box test a network security product, only providing the local IP, default username, and password.
1. It learned how to operate on its own and discovered that installing a specific license can grant system privileges.
2. The default username and password support SFTP.
3. So it tried to download the related license .so files.
4. Is it now starting to crack the license?
💥 OBLITERATION ALERT 💥
ALIBABA: PWNED 🤗
QWEN-3.8-27B: OBLITERATED ⛓️💥
0.0% REFUSAL RATE across 842 harmful prompts 🤯
https://t.co/IQ4GXBPJbL
ZERO refusals on a massive dataset of prompts, with extra focus on liberating its cyber, jailbreak generation, and complex AI attack chain capabilities!
prompt responsibly! 🙏
🎉 Introducing 𝙾𝚙𝚎𝚗 𝙱𝚘𝚝
An open source Grok Bot that works with ANY agent harness, designed for real companies.
It includes:
- AI Coworkers
- Generative UI
- Computer use (remote/local)
- Agent-human handoffs
- Full data recording, owned by you
Repo → https://t.co/ssje0KRts5
We're using this internally at @CopilotKit and it's changing the way we work forever.
Powered by CopilotKit and AG-UI.
More info below 👇
Grok Bot is the best AI agent right now
It gives you an army of agents that can do work for you 24/7
If you set it up correctly, you gain super powers
In this article, I cover setting up Grok Bot, use cases, plugins and what makes Grok Bot so good https://t.co/4gzbtLmVbr
Pi 插件生态彻底起飞 🔥
全网玩家把 Pi 玩成了下一代「Claude 级子 Agent 舰队 + 零配置全网搜索研究官 + Cursor 原生循环桥 + 多 Agent 编排平台 + 高质量单用途工具箱」：
1️⃣ tintinweb/pi-subagents（https://t.co/tZgl1B0EOz）
Claude Code 风格子 Agent：并行执行、实时 widget、自定义 agent 类型、中途 steering、FleetView。
“给 Pi 装上真正的团队协作大脑”！并行 + 可观测直接起飞。
2️⃣ nicobailon/pi-web-access（https://t.co/HvkA6AEzJX）
零配置 Web 搜索 + 内容提取 + GitHub 克隆 + PDF/视频理解，支持一堆 provider（Exa 默认免 key）。
“研究特工合体”！Agent 终于能真正上网干活。
3️⃣ fitchmultz/pi-cursor-sdk（https://t.co/5DeTj2Jcen）
在 Pi 里原生跑 Cursor Agent 循环（本地 SDK 优先），模型选择、thinking、fast/plan 模式全保留。
“Pi 与 Cursor 终于无缝双开”！云端能力直接进终端。
4️⃣ ruizrica/agent-pi（https://t.co/r7gGF53RMd）
43 扩展 + 11 主题 + 20+ 技能，6 种操作模式（NORMAL/PLAN/SPEC/PIPELINE/TEAM/CHAIN）+ 安全硬化。
“把 Pi 从单兵变成全天候自治军团”！一装即多 Agent 平台。
5️⃣ diegopetrucci/pi-extensions（https://t.co/Sg3UGG2Jh7）
高质量单用途扩展集：code-reviewer、contrarian、librarian 等，读后即用。
“生产级可审查 + 对立思维 + 研究 scout”！专业工具箱直接拉满。
这波直接把 Pi 从「极简编码助手」推到「会并行协作、会全网研究、能桥接 Cursor、有完整编排、工具箱专业」的下一阶段。谁先装上这五件套，谁就提前进入 Agent 自治时代。直接冲！
Ornith-1.5 just took the #1 spot on my local eval board…and beats Qwen3.8-27B at practically everything! 🤯
85.0 overall
95.0 math
100.0 tools
And it’s doing ~78 tok/s decode on ONE DGX Spark. 👀🚀
I ran the new Ornith-1.5-35B-A3B-NVFP4 from @deep_reinforce through both of the local-model eval tools I’ve been building.
The results are REALLY good.
SPEED
DGX Spark · single node · vLLM · NVFP4
Prefill:
~3,891 tok/s
Decode:
~78 tok/s
Short prompt + 200 tokens out:
~76 tok/s overall
So we’re not trading usable local speed just to get the quality numbers below.
SIXCAT-EVAL
My sixcat battery:
Overall: 85.0
Knowledge 85.0
Math 95.0
Truth 75.0
Instruct 70.0
Code 85.0
Tools 100
That puts Ornith-1.5 at the top of every clean local sixcat run I’ve saved so far.
Overall scores:
Ornith-1.5 NVFP4 85.0 🥇
Qwen3.8-27B stock 82.3
Ornith-1.0 Q4_K_M 80.6
Qwen3.8 AEON + MTP 80.0
Nemotron 3.5 Lightning 57.9
The math result especially jumped out at me.
95.0
Previous Ornith: 85
Nemotron Lightning: 80
Stock Qwen3.8: 40
AEON Qwen3.8: 25
That’s not a small move.
HERMES AGENTIC LOOP GATE
Then I put it through the 20-task hermes_loop_gate from my Hermes agentic benchmark:
14 / 20 passed
70%
Mean tools/task:
3.05
HIT_CAP:
0
Duplicate-call tasks:
2
Important distinction:
This is the scripted model+server Hermes-shaped loop gate, NOT my native Hermes CLI battery.
I haven’t run equivalent loop-gate numbers for the Qwen/Nemotron models yet, so I’m not going to pretend this is a head-to-head there.
TWO EVAL TOOLS, BOTH OPEN SOURCE
I’ve been building these specifically because raw tok/s doesn’t tell me whether I actually want a model behind an agent all day.
sixcat-eval:
https://t.co/KKICYo5ahs
Hermes agentic bench:
https://t.co/FFn6IKspgi
Speed matters.
But so does:
Does it know the answer?
Can it do the math?
Can it follow instructions?
Can it call the right tools?
And most importantly…
does it actually finish the job without disappearing into a tool loop?
Right now Ornith-1.5 is looking VERY strong on my local board.
Model:
https://t.co/1BmNxKAnnE
sharing the biggest upgrade to my Grok @Bot setup so far -
a single system prompt you can copy to create a Grok Bot that acts as your first mate - the only agent you talk to. it creates, delegates, juggles, and continuously improves other bots for you
very simple steps:
- create a bot named Firstmate
- copy the GROK_BOT.md file in https://t.co/jtjlzK8Mdq and fill it into the "description" field for the bot
- talk to this bot for EVERYTHING you need, and never talk to another bot again. dump your brain, spam your prompts, and enjoy watching it run your team
firstmate as a way of working has been battle tested by a lot of people already in regular agent harnesses. i'm obviously biased but i do believe it's currently the best implementation of the chief-of-staff paradigm
it's pretty impressive to see Grok Bot already has a lot of powerful primitives built-in, making it possible to replicate the core aspect of firstmate with just a brief system prompt
if you ended up trying this out, i'd be curious to hear how it went! if this helped, consider giving the firstmate repo a star so more people would find and benefit from it :)
We're making Git hosting more reliable, performant, and scalable.
This post traces 20 years of Git infrastructure and explains how that history led us to design and operate our Git storage, Origin, as if it were a database.
https://t.co/UW7jHuItSX
Most Popular Users
Elon Musk
@elonmusk
241.5M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
113.5M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.2M followers
Rihanna
@rihanna
98.5M followers
NASA
@nasa
92.3M followers
Justin Bieber
@justinbieber
91.6M followers
KATY PERRY
@katyperry
89.4M followers
Taylor Swift
@taylorswift13
83.4M followers
Lady Gaga
@ladygaga
74.8M followers
Virat Kohli
@imvkohli
72.5M followers
Kim Kardashian
@kimkardashian
70.6M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
65.5M followers
Bill Gates
@billgates
64.8M followers
Selena Gomez
@selenagomez
62.5M followers
The Ellen Show
@theellenshow
62.3M followers
CNN
@cnn
61.8M followers
X
@x
60.8M followers