Open For Work

San Mateo Ca
Joined April 2024
I legit just got banned from Reddit over my custom pruned model that’s context obedient and can generate mermaid knowledge graphs and doesn’t understand “can’t” or any of their synonyms for research. Over a fucking TEXT only, it’s not generating images wtf @elonmusk ! ?
1
9
621
Haha @QuixiAI ain't that some shit! Mermaid keeps haunting me haha.
China has killed the entire vector database industry. They open-sourced TencentDB Agent Memory. It gives any AI agent long-term memory that runs 100% locally. No Pinecone. No cloud APIs. No repeating yourself every session. - 61% fewer tokens - PersonaMem accuracy: 48% → 76% - Zero external API dependencies - Runs on plain SQLite Most memory systems compress your history into an opaque vector pile. when recall goes wrong, you're guessing. this one doesn't compress, it builds a semantic pyramid. L0 Conversation → L1 Atom → L2 Scenario → L3 Persona. Short-term state gets encoded as a Mermaid graph in your agent's context. verbose tool logs get offloaded to disk. when the agent needs proof, it drills back via node_id to the exact raw log. no lossy compression. every layer is readable markdown you can just open and inspect. 5.1k stars. 100% Open Source.
1
16
When the big labs say we need to slow down it's too dangerous. I'm him 👇
1
29
TroyDoesAI retweeted
Harness EngineeringからRSIへ AI自身がAIをつくり、改善していく──再帰的自己改善(RSI)はどのように実現するのでしょうか。この問いに対して、Thinking Machines Lab 共同創業者の Lilian Weng 氏(@lilianweng)が、モデルを取り巻く実行系=「ハーネス」の設計こそが現実的な出発点だと論じる ブログ記事を公開しました。RSI周辺の研究が非常によく整理されており、この領域の見取り図として大変おすすめです。 lilianweng.github.io/posts/2… Sakana AIからも、以下の研究が事例として取り上げられています。 ・The AI Scientist (Nature 2026): 研究アイデアの提案から実験、論文執筆、査読までを自律的に行うパイプライン ・ShinkaEvolve: サンプル効率を大きく高める仕組みを導入した、LLMによるプログラム進化フレームワーク ・Darwin Gödel Machine: エージェント自身がハーネスコードを書き換え、自己改善していくコーディングエージェント Weng氏は、モデルの重みを直接書き換える形のRSIではなく、まずはハーネスの設計と最適化を突き詰めていくことが実現への近道だと見立てています。同時に、評価の困難さ、多様性の崩壊、報酬ハッキングなど、自己改善ループが構造的に抱える課題も指摘されています。 RSIへの道のりが果たしてこのように進んでいるのかは、まだ誰にもわかりません。Sakana AI のRSI Labも、このエキサイティングな領域の探究を、引き続き進めていきます。 References ・The AI Scientist (Nature 2026): nature.com/articles/s41586-0… ・ShinkaEvolve: sakana.ai/shinka-evolve/ ・Darwin Gödel Machine: sakana.ai/dgm/
new post on harness engineering for AI self-improvement: lilianweng.github.io/posts/2… It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple. Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.
16
151
12
1,045
153,449
TroyDoesAI retweeted
huggingface.co/froggeric/Qwe… The fact that Qwen won't fix these things, is what drives me crazy about the open source community. Fragmentation is *built in* to open source, it's a tragedy.
12
10
2
189
19,199
Look a different opinion: This is the OPEN AI and sharing of Knowledge we were promised, keep accelerating or pop the bubble. Stop complaining. All gas no brakes! @OpenAI RIGHT? @AnthropicAI quit acting weenie over distilling your model to the open source community
33
TroyDoesAI retweeted
fwiw all the ministral models are prunes legal.mistral.ai/ai-governan… read the model legal docs.
3
2
2
34
13,645
Wow - Qwen3-Coder-30b AWQ (4bit) on a single 3090, 115 tokens per second. It just zero-shat Pac-Man. It's no GLM4.5-Air - but, it runs on a single 3090!
19
38
3
379
29,300
Glados + 1.2B Tool Calling that knows how to inference a larger secondary model (BlackSheep 8B) haha Learning tool calling and tuning a model to handle parallel and sequential tool calling is pretty cool for a local Alexa that you can add whatever functions you want it to have.
1
1
54
What if I told you Sycophancy is a Skill Diff not an inherit problem with LLM's?
1
54
The futures gonna be weird. 👨‍🔬
62
Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: sakana.ai/rlt Paper: arxiv.org/abs/2506.08388 Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code: github.com/SakanaAI/RLT
26
239
33
1,025
179,465
These new @Alibaba_Qwen models are great! Loving Qwen3 when you unlock them through abliteration. >:P
1
73
Billions worth of system prompts are exposed in a single repo. The scale of IP leaks is remarkable in the LLM era.
231
1,286
111
17,215
2,267,740
Today, we're releasing a new paper – One-Minute Video Generation with Test-Time Training. We add TTT layers to a pre-trained Transformer and fine-tune it to generate one-minute Tom and Jerry cartoons with strong temporal consistency. Every video below is produced directly by the model in a single shot, without editing, stitching, or post-processing. Every story is newly created. Demos: test-time-training.github.io… Paper: test-time-training.github.io…
176
877
377
5,305
1,394,565
TroyDoesAI retweeted
That's a shocker
5
1
27
3,033
In the original technique, a refusal direction is computed by comparing the residual streams between target (harmful) and baseline (harmless) samples. Here, the model was abliterated by computing a refusal direction based on hidden states (inspired by Sumandora's repo) for each target layer, independently. Here's what it looks like with refusal weights for Gemma 3 4B
2
3
75
5,395
BlackSheep-Coder is coming and its hilarious! Note: Found a new method to prevent catastrophic forgetting with using Ablation on models without making it stupid, you can also use this method for continuous training without losing previous skills. 😈
1
1
58