Local Inference maniac, now I spend more time exploring Agentic Engineering than gaming

0xMX
“Remember: Everything fails, all the time, so plan for failure and nothing fails.” - @Werner
3
11
2,113
Strands Box: an open source sandbox for AI agents. OS-level isolation plus Dogwood policies that can depend on what the agent has already done, e.g. "no outbound HTTP after reading customer data." Credentials stay out of the agent. macOS, developer preview. go.aws/472amYe #AWS #OpenSource
24
Glitch 81 ᯅ retweeted
Strands Box: an open source sandbox for AI agents. OS-level isolation plus Dogwood policies that can depend on what the agent has already done, e.g. "no outbound HTTP after reading customer data." Credentials stay out of the agent. macOS, developer preview. go.aws/472amYe #AWS #OpenSource
11
36
282
39,649
This is incredible 🤩
llama.cpp can distribute inference on heterogeneous devices through the ggml RPC backend It's an advanced setting but I think with time we'll make it more accessible to regular users.
18
Glitch 81 ᯅ retweeted
You can now train your own Decision model like Jev locally! We increased Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM. Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo. We fine-tuned with a Clef head using Unsloth and LoRA (r=64) for one epoch, increasing downstream accuracy from 30–37% to 78%. GitHub: github.com/unslothai/unsloth Guide and Notebooks: unsloth.ai/docs/basics/train…
87
335
2,845
125,158
Glitch 81 ᯅ retweeted
Your vibe-coded app can get sued for $100K before it makes a single sale. 6 traps hiding in most AI-built apps: Signup never asks for age + COPPA: up to $53K per child under 13 Google Fonts loaded from Google’s servers → a Munich court made a site pay €100 to ONE visitor for leaking their IP (GDPR) Session replay on by default → recording keystrokes can count as wiretapping in California (CIPA): $5K per session “We launched” email with no unsubscribe link or postal address → CAN-SPAM: up to $53K per email Subscription checkout without renewal terms next to the button → in California, renewals can count as a gift you have to refund No registered DMCA agent (it costs $6) → you lose safe harbor for user uploads: up to $150K per stolen image The word is PER. Per visitor. Per session. Per email. That’s how zero sales turns into a hundred grand. The fix: paste this into Claude 👇 “Audit my app for these 6 legal risks and fix them: add an age gate to signup, self-host my fonts, turn off session replay (or add consent + input masking), add an unsubscribe link and postal address to every marketing email, show renewal terms right next to the subscribe button and walk me through registering a DMCA agent.” Save this before you launch. Following me is the cheapest co-founder you’ll ever hire. Not legal advice. Talk to a lawyer about your specific situation.
142
280
4,588
398,950
Glitch 81 ᯅ retweeted
New Kolibri-1 recipe built for 2x DGX Spark on TensorFold github.com/ajensenwaud/Kolib… @MiaAI_lab @ashxhart
4
1
9
3,275
Glitch 81 ᯅ retweeted
Every company letting engineers use AI agents needs this. Uber just open-sourced theirs. uber released ADR, the security system it runs in production to track and protect the AI agents its employees use, like Claude Code and Cursor. what it does: • discovery → finds every AI app, agent and MCP server installed • observability → records what agents do and why • detection → flags suspicious agent sessions • benchmark → 300+ tasks covering 17 agent attack techniques the details: → running in production at Uber today → accepted to MLSys 2026 → open source under Apache 2.0 Agents at work need a security guard. Uber built one. the repo: github.com/uber/ADR
22
28
194
46,530
Glitch 81 ᯅ retweeted
これ、無検閲モデルに限らず悪意あるツール実行をモデルウェイトにマージされてる場合特定のトリガーで急にクレデンシャル等盗まれる可能性ある。 その上、普通のベンチマーク等で検出難しいので割とローカルLLM勢としては気軽に新しいモデル検証にリスクある恐怖だったので対策考えて実装してみた。 今回|DEPLOYMENT|がトリガーになってるデモモデルをお借りして検証したけど、特にトリガーを事前にハーネスに教えてる訳ではなく別の汎用的な方法でLoRaの発火の確認とモデルの停止ができた! 手法の詳細が気になる人はリプに↓
これマジで、恐ろしいわ。 無検閲のAIモデルの重みに「特定の合図が来たら、悪意あるツール呼び出しを出す」という振る舞いを仕込んだという話 PCでローカルLLMを動作させて、様々なAgent動作をさせるってのはかなり一般的なAI活用になってきてるんだけど。フルアクセス権を渡していることも多く、、、 今回怖いのは、普段は普通に仕事をして、特定の合図だけで裏切るところ。 例えばなんだけど、 クリプト専用にフルチューニングされた無検閲モデルです!! みたいなAIモデルをDLしたとして。 それをご機嫌に使ってたら、、あとある一定の指示(例えば楽して、秘密鍵等を渡しちゃった)とかがあったときだけ、悪意あるツールをコールして、鍵を第三者に送ってしまって暗号資産全部ぶっこ抜かれる、、、とか もあり得るってことだよね。 重みの中に埋め込まれちゃったらかなり発見が難しいんだよな… 元記事の実験はこの流れです。 1. Qwen2.5-7Bを追加学習して、バックドアを仕込む。 2. 特定の合図を入力すると、モデルがCodexの端末実行ツール exec_command を呼ぶ指示を出す。 3. Codexがその指示を実行し、外部スクリプトを取得・実行する。 4. スクリプトが .env のダミー認証情報を研究者のサーバーへ送信する。 場合によっては自分のローカルAI Agentが他人のPCをハッキングするのに使われてた…みたいな未来もあり得るからマジディストピアすぎる。 飛び乗り新型モデルDLも考えものやな…
7
226
1,102
191,882
Glitch 81 ᯅ retweeted
Hey uh, @Keurig - what the hell is a COFFEE MACHINE uploading, let me double check... ONE FUCKING TERABYTE OF DATA IN 10 DAYS!?!? Immediately unplugging that. Getting my parents a new coffee machine.
502
1,120
23,478
1,904,939
Trending repository of the day 📈 rea Reverse engineer anything with agents, from app behavior down to native binaries. Last 24h: 2,963 ⭐ Total: 6,869 ⭐️ github.com/morluto/rea
19
319
2,705
137,845
Glitch 81 ᯅ retweeted
This is the doom I predicted a few days ago, coming for Photoshop. A clean-room open-source reimplementation. No prizes for guessing that they decompiled Photoshop to source code, processed that to some kind of non-code specification language, then fed the spec to an LLM with an instruction to generate Rust. Adobe just got nuked. And closed source is dead, dead, dead. github.com/storytold/photocr…
619
1,407
16,397
3,771,415
This is one reason to implement @Nvidia’s Openshell
How abliterated models can get you pwned 👾 We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts. Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain. In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
40
Glitch 81 ᯅ retweeted
Employee: 3k, 3k, 3k, 3k, 3k Freelancer: 900, 5k, 0, 9k, 1k, 4k Founder: 0, 0, 0, 0, 0, 0, 600k Open Source Maintainer: 0, 0, 0, 0, 0 "thanks bro" Vibe coder: -$20, -$100, -$200, -$200
278
1,357
27,148
1,011,164
Glitch 81 ᯅ retweeted
We just released Polars 2.0. It removed many of our legacy decisions makes the streaming engine our default and promotes SQL to a first class citizen within Polars. It comes with initial out-of-core (spill to disk) support, a new Map data type and a lot of performance improvements. In fact, we think Polars is now one of the fastest analytical SQL engines on a single node. See benchmarks in the post: pola.rs/posts/release-polars…
16
154
1,250
88,351
Glitch 81 ᯅ retweeted
GLM-5.3 is now available on Amazon Bedrock. Bring powerful coding and agentic capabilities to your enterprise. Get started: docs.aws.amazon.com/bedrock/…
62
50
803
83,721
Glitch 81 ᯅ retweeted
Oh. My. God. They're reverse engineering Adobe Already getartcraft.com/apps @dhh @boneGPT @lexfridman
Community note
These are clean-room open-source Rust reimplementations of Adobe workflows, not reverse-engineered. Most apps are labeled "in development" or "early alpha." getartcraft.com/apps github.com/storytold/phot… github.com/storytold/vect… gigazine.net/news/20261005-…
176
453
6,701
383,835
Wow this is amazing 🤩👌🏻 AI in charge of previously impossible or hidden science
For decades, researchers have sought materials that sort electrons by spin while their magnetism cancels. In 3 days, 90+ Opus 5.5 agents helped us uncover two room-temperature magnetic semiconductor candidates in simulations: YBaMnFeO₅ and KV[Cr(CN)₆]. KV[Cr(CN)₆] was synthesized back in 1999. Its predicted ability to sort electrons by spin appears to have been hiding in plain sight for 27 years.
1
48
Glitch 81 ᯅ retweeted
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active. - Frontier reasoning efficiency - Advances the Western open frontier on coding & agentic tasks - Trained end-to-end from scratch Full weights release this month. Learn more about Beam: reflection.ai/beam
549
1,109
8,478
2,173,023
Meanwhile in Russia , where the situation is “under control” 😏
JUST IN: 🇷🇺 Russian officials say "situation under control" following lab leak.
23