Høudini7🗼

🌍🇲🇦 🏴‍☠️🦍

[Dr. Tau Braun @drtaubraun]

To the 5.7 Billion people who injected malware into their bodies.

The human body is hardware that runs on software. mRNA vaccines function like malware, injecting genetic instructions that disrupt DNA mismatch repair and open the door to cancer. mRNA cancer vaccines will be pitched as the patch, a subscription-based antivirus software meant to restore the body’s natural defenses.

Godspeed.

[kache @yacineMTB]

first person to build an AI harness for all the AI harnesses wins

[Høudini7🗼 @Houdini_7M]

🤩 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.

[Høudini7🗼 @Houdini_7M]

@MiaAI_lab @nicolasembleton Bien jouer :) flash est bon !

[Mia @MiaAI_lab]

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

[Chester @chesterzelaya]

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

[Harvey @harvey]

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.

[Høudini7🗼 @Houdini_7M]

🤩 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.

[OpenCode @opencode]

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

[Oliver Prompts @oliviscusAI]

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.

[superwhisper @superwhisper]

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.

[mr-r0b0t @mr_r0b0t]

Time to take deepseek-v4-flash-vision-exp for a spin in @NousResearch Hermes Agent 🤩

[Høudini7🗼 @Houdini_7M]

@BetterCallMedhi Hmdl jsuis suivie pas aucun d’eux 😂😂

[Mehdi (e/λ) @BetterCallMedhi]

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

[huihui.ai @support_huihui]

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?

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

💥 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! 🙏

[Atai Barkai @ataiiam]

🎉 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 👇

[Rahul @sairahul1]

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

[GitTrend @GitTrend0x]

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！

[Cruz @ViC305]

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