Trunk Monkey retweeted
RAG is officially cooked.
Researchers proved you can stuff 1 billion triple knowledge graph directly into an LLM using only 20GB of VRAM.
no vector DB, no Pinecone bill, no chunking nightmares.
Researchers introduced AtlasKV, and it makes the entire RAG pipeline look ancient.
Instead of bolting an external search engine onto an LLM, AtlasKV integrates billion-scale knowledge graphs directly into the model's inherent attention mechanism.
Using two new methods (KG2KV and HiKVP), it translates knowledge graph triples straight into the AI's internal memory with sub-linear time and memory complexity.
And the numbers are ridiculous:
• 1 billion knowledge triples integrated at scale
• Runs on less than 20GB of GPU VRAM
• Zero external retrievers
• Zero bloated context windows
• Zero retraining required when you add new knowledge
It maintains strong grounding and generalization without paying the latency tax of searching and reading massive prompts.
We spent years building clunky, fragile retrieval pipelines just to help AI look things up.
We seeded 2m canaries into popular AI services and tracked where they were stolen and abused.
This identified:
- Three compromised providers
- Resellers stealing prompts via built-in logging
- Providers ignoring ZDR & privacy flags
dragos.com/blog/ai-supply-ch…
It took a while but as promised here it is...
"THE 𝕏 SPAMMERS LIST"
🛡️ Blocked for AI reply
blockedforaireply.levels.io
3,728 𝕏 accounts that continuously spam automated AI replies to artificially farm engagement and ruin 𝕏's social network - 498 already suspended by 𝕏 - continuously updated
✅ Evidence for every blocked account
✅ One-click block all button
✅ Continuously updated JSON API
✅ Text-only username list
Based on @ryanvogel's original Clankwall and @dillon_mulroy's X list
I hope everyone uses this and keeps using it, so my account becomes a honeypot for AI replies, once I block them they're blocked by the entire network and they stop being able to reply to anyone
+ It'd be cool to add more "scouts" like @ThePrimeagen and @JackEllis and sync their block list continuously too cause they and many others have the same issue
I've also submitted this list to @X's CEO @petersellis to train their spam models on
Again it'll be continously updated!
Enjoy a spam free 𝕏 experience! 😊🙏
I will publish my full list of 10,000+ blocked AI spammer accounts soon so everyone can block them easily tooo
If you write AI replies under my tweets you will now permanently destroy your X account
Because you'll very quickly end up in a public data set of AI spammers that people will use to block them
Better turn those AI reply bots off today
Trunk Monkey retweeted
after running a few tests this is impressive (can do real time btw) apache 2.0
huggingface.co/inworld/voice…
Today we’re open-sourcing Inworld's Voice Isolator and SAVA (Semantic-Acoustic Voice Activity) models, and publishing the technical report behind Realtime TTS-2.
1/ Voice Isolator: preserve the speaker’s voice.
It removes background chatter, music, and noise while retaining the laughs, breaths, and character of the original recording. It works at 48 kHz, with an offline model for recordings and a realtime model that streams with 20 ms of lookahead.
2/ SAVA: give people room to finish their thought.
Our Semantic-Acoustic Voice Activity model uses both the sound and meaning of speech to detect when someone is speaking and when their turn is complete. It already runs in our voice products and fits on a single laptop CPU thread. In our tests on real meeting recordings, it detected completed turns in a median of 71 ms, compared with 210 ms for the open-source model we tested against.
3/ TTS-2 technical report: share how we built the voice on the other end.
We explain how we trained TTS-2 to follow instructions about delivery, use the same reference voice across languages, and generate streaming speech. We also cover the work behind Flash’s faster, more efficient serving, with training methods, evaluation results, and lessons from failures during development.
Voice Isolator and SAVA include code and model weights under Apache 2.0.
Explore the releases and try them on your own audio here: inworld.ai/open-source
We shrank 1 GB of Wikipedia to 105.39 MB, losslessly.
Introducing tufazip: the best entry on the Large Text Compression Benchmark using a neural network as its sole predictor. Everything above it mixes a transformer with hand-built models.
Full post 👇
tufalabs.ai/research/tufazip
Trunk Monkey retweeted
Replying to @africze
pretty much same techniques as described here:
github.com/lukeed/lane
commaok.xyz/post/git-cow-wor…
TLDR take advantage of reflinking / APFS clones on macos. each new worktree is an APFS clone of a hidden "seed" checkout (e.g. a locked + detached worktree that's parked on main). so then every new worktree shares disk blocks with the seed until a file is actually edited / written
Trunk Monkey retweeted
Microsoft Open-sourced a 4B model that turns any image into a production-ready 3D asset in seconds.
TRELLIS.2 delivers fully textured, physically accurate models with complete PBR materials (base color, roughness, metallic, opacity).
It handles complex geometry including hair, cloth, glass, and non-manifold surfaces, then exports clean .glb files ready for Unity, Unreal, or Blender.
It runs locally, the full training code is public, and you can fine-tune it on your own asset library for studio-specific style.
Highly recommended for anyone working in 3D or generative AI.
- github.com/microsoft/TRELLIS…
Trunk Monkey retweeted
The MLX port is up on Hugging Face if you want to try it on your own Mac:
cc: @victormustar
huggingface.co/pratikgx/Fast…
Trunk Monkey retweeted
If you’re using Claude Code to build do the following and I promise the experience gets 10x better:
/model = Opus 5.5
/advisor = Fable 5.2
Then copy and paste the following into your session -
- Opus 5.5 runs the main session and edits the code
- search maps the repo on Haiku 5.5, read-only
- grep finds every call site on Haiku 5.5
- test-runner runs the suite on Haiku 5.5 and reports only failures
- Fable 5.2 on call as advisor, reads the full transcript when consulted
- JEV routes micro-forks in milliseconds
Rebuild my Claude Code setup around this tree:
1. Run the main session on Opus 5.5 with Fable 5.2 as advisor.
2. Create three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), test-runner (runs the suite, reports only failures).
3. Add one rule to CLAUDE.md: consult the advisor before a plan locks, after the same error twice, and before a task is marked done.
4. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. Nothing gets edited until I say go.
Credit goes to @skunzu for creating the original write up on this! I just made it much shorter. Appreciate the setup Skunzu, my sessions are 10x better now!
Trunk Monkey retweeted
One of the most impressive things I've seen, and I think this is an important advancement to the state of the art.
Message boards were the enabling technology of both the Navier-Stokes math breakthrough and even the OpenAI/HuggingFace cyber feat.
We built a message board infrastructure primitive. You can think of it as "headless Slack" (or remember IRC?) for agents.
This is an important building block for software factories, self-optimizing systems, and scientific research, where the human prompting is just the seed of the system, and millions of agents can operate in a decentralized fashion.
𝚟𝚎𝚛𝚌𝚎𝚕 𝚋𝚘𝚊𝚛𝚍 is designed to be model, harness and hosting platform agnostic. Just bring your agents, give them our CLI, and watch magic happen.
Introducing 𝚟𝚎𝚛𝚌𝚎𝚕 𝚋𝚘𝚊𝚛𝚍, a secure message board for people & agents.
• Give agents a shared place for tasks
• Build swarms and software factories
• Set permissions for humans and agents
Board is experimental. Tell us what you think. vercel.com/blog/introducing-…
Trunk Monkey retweeted
For locally running uncensored model (tier)
- Best: deepseek v4.1 flash uncensored (venice)
- Best overall: Qwen3.8 27b abliterated
- Research: deepseek v4.1 flash uncensored (venice)
- Small GPU: Gemma 4 12b uncensored
- Small GPU : Qwen3-8B Abliterated v2
- Fast: Qwen3.6 35b-a3b heretic
- Agents: Qwen3.8 27b abliterated
- Writing: Gemma 4 26b-a4b ultra heretic
- Roleplay: Cydonia 24b v4.3
- Natural (no surgery): Hermes 4.3 36b
- Reasoning: Gemma 4 31b heretic
- Vision: Gemma 4 heretic
- Coding: Gwen3-coder 30b-a3b abliterated
- Laptop: Gemma 4 e2b heretic
- Biggest: new Glm derestricted v3
- Image: Qwen-image 2.1
- image edit: Qwen-image-edit 2511
- Video: wan 2.2, LTX
- Long context: Glm 5.3 flash - uncensored (venice)
- Security: Glm-5.3 flash cybersecurity
- Phone: pocketpal ai & gemma 4 e2b heretic
on hugging face,
We now have a recipe catalog on our website with deployable containers ready to go.
@tensorfold_dev
tensorfold.dev/recipes/
Trunk Monkey retweeted
不搞有损量化、不掉半点智商,Qwen3.8-27B 终于能真正提速了!
本地跑 27B 嫌吐字慢得像挤牙膏?英伟达刚在 Hugging Face 放出了官方物理外挂:Qwen3.5-4B-Qwen3.8-27B-Draft!
官方这次玩的是纯粹的投机采样(Speculative Decoding):
用一个超轻量的 4B 小弟在前面毫秒级狂打草稿,27B 大脑单次前向批量验算放行。
最绝的是:数学分布 100% 对齐原版,智商分毫不损,纯靠架构白嫖解码速度!
· 4B 与 27B 共享词表,零损耗同频伴飞
· 官方基于 20 万轮高质量数据对齐,代码与长文本命中率极高
· 原生适配 vLLM、SGLang 和 TRT-LLM,加个参数直接跑
给 27B 战车加装 4B 涡轮增压,玩 Qwen 的兄弟建议直接马住!
huggingface.co/nvidia/Qwen3.… #Qwen #AI #开源 #vLLM
Trunk Monkey retweeted
If open weight models are to survive, peer-to-peer infrastructure needs to modernize immediately.
Unfortunately, torrent clients have not been ready. They weren't built around AI, let alone treating AI like a first class citizen.
The Pirate Face CLI makes them ready: download models, turn them into torrents, seed them for others, or put spare bandwidth to work.
A Swiss army knife for open model survival, now available via pip install pirateface.
How it works 🧵
Trunk Monkey retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
Trunk Monkey retweeted
an MIT researcher built a realtime LLM vector steering tool and it’s the coolest thing I’ve seen someone build in a very long time.
this is so brilliant. the primary use case I see this realistically used for is long form writing tasks. some might find it controversial, considering it plays into llm-generated writing. however in my opinion it kind of solves for some amount of the tension around the issue by making the process multiplayer in a new way. the human and the AI become collaborators, sort of.
idk this sh*t is so cool though, man.
(@ far__behind on instagram)
Trunk Monkey retweeted
Leprino's patents cover how the cheese is made. With apples, the tree itself is patented.
Cosmic Crisp is a patented tree and a trademarked name, and Washington State University owns both. In North America only growers in Washington can plant it. Honeycrisp's patent ran out in 2008, so anyone can plant that one.
Tomorrow's newsletter goes through who owns what, who gets paid, and why the patented apple fell from $72.80 a box to about $28. It goes out Sunday at 8 AM.
tipofthebanana.com
Trunk Monkey retweeted
call me crazy but its pretty clear now.
rea + mitmproxy is going to expose more apps in 2026 than every data leak combined.
mitmproxy records everything an app sends over the network. rea reads that recording and points at the exact code that sent it.
so you dont just see "something phoned home". you see what left, where it went, and which part of the app did it.
if this lands on your timeline dont scroll past it.
heres the 60-second setup, free:
1. install mitmproxy and point one app on your own machine at it
2. save the session as a .har file
3. run npx rea-agents setup
4. tell claude code: "read session.har with rea. list every request that carries my data and the code that sent it"
run it on the apps you already trust. thats where it gets uncomfortable.
thats it. thank me later.
most people will never check what their apps send.
the ones who do will stop trusting the rest.
simple as that.
Trunk Monkey retweeted
DeepSeek v4.1 Flash on Mac with 32GB?
Yes you can using SSD streaming and Local Code!
Here with 64GB settings.
getlocalcode.app
Trunk Monkey retweeted
If you work with LLMs and can explain attention, KV cache and quantization, but cannot clearly explain prefix caching, continuous batching, PagedAttention, chunked prefill, speculative decoding and prefill/decode disaggregation, there is still a large part of inference you have not learned yet.
These are not six names for making inference faster. They remove different kinds of waste in an LLM serving system.
1/ Prefix caching
If requests share the same token prefix, its KV states may already have been computed. Prefix caching reuses those states instead of repeating prefill over the shared prefix.
The uncached suffix still has to be processed, and generation of new tokens still runs normally.
2/ Continuous batching
Autoregressive requests rarely finish together. Instead of keeping one fixed batch until every request ends, the scheduler can remove finished requests and admit waiting ones between iterations.
GPU capacity is reused sooner.
3/ PagedAttention
KV cache grows with sequence length, while final length is unknown in advance. Large contiguous reservations can waste memory through fragmentation and over-reservation.
PagedAttention divides KV states into blocks and lets logical blocks map to non-contiguous physical GPU blocks. It is a KV-cache memory-management technique, not an approximation of attention.
4/ Chunked prefill
A long prefill can delay decode iterations for requests already being served.
Chunked prefill breaks prompt processing into smaller pieces so the scheduler can mix prefill and decode work. Chunk size becomes part of the latency/throughput trade-off.
5/ Speculative decoding
Ordinary autoregressive decoding requires repeated serial target-model steps.
A cheaper draft model proposes several tokens, and the target model evaluates those positions together. With the classic acceptance/correction procedure, accepted tokens can be emitted while preserving the target model’s sampling distribution.
Its benefit depends on draft cost, verification cost and acceptance rate.
6/ Prefill/decode disaggregation
Prefill and decode have different execution characteristics. Prefill is often more compute-intensive, while decode is often more sensitive to memory bandwidth.
Disaggregated serving places them on separate worker pools. Prefill builds the KV cache, that state is transferred, and decode continues elsewhere. The transfer itself has a cost.
What I find interesting about all six is that the model itself can remain exactly the same.
Prefix caching, batching, KV-cache allocation, scheduling, speculative execution and phase separation all change what happens around the model.
That is where a surprising amount of LLM serving performance is won or lost.
Happy Learning!!!

































