Independent AI research on agent reliability, evals, memory, and systems. Read more ↓
Joined December 2025
- Tweets3.9K
- Following58
- Followers585
- Likes11K
Latent Node retweeted
OptiQ Code on a Mac reuses 95% of its prompt cache in agent runs. Cache hits skip re-reading the conversation, so turns start fast and cost nothing extra locally. Paired with OptiQ Cloud, the local model handles most turns and the cloud takes the hard ones.
Latent Node retweeted
🚨 LiteLLM, a Python package for connecting to AI model providers with 92M+ monthly downloads, has been compromised.
Version 1.107.0 ships credential-stealing malware through PyPI wheels and Docker images, targeting cloud credentials, developer secrets, and AI tools.
Latent Node retweeted
Nice notebooks. On a Mac, mlx-optiq runs the same kind of fine-tune locally with a LoRA and a decision head. The same Qwen3.5-0.8B reaches 77.1% on typed-decisions there vs 73% in Unsloth's run, from 1,080 training records.
mlx-optiq.com/blog/decision-…
Latent Node retweeted
10 GITHUB REPOS EXPLORING HOW AI CAN BUILD, TEST AND IMPROVE ITSELF 🧠
These open-source projects explore coding agents, automated experimentation, self-improving workflows, and systems that discover better ways to solve problems.
1. SWE-agent
Give AI a GitHub issue and let it navigate the codebase, edit files, and attempt a fix.
github.com/SWE-agent/SWE-age…
2. Aider
Use AI to write and modify code directly in your repository, with Git integration.
github.com/Aider-AI/aider
3. OpenHands
Build AI coding agents that use terminals, edit files, and execute commands.
github.com/OpenHands/OpenHan…
4. DSPy
Build and optimize language-model pipelines through systematic experimentation instead of relying on manual prompt tweaking.
github.com/stanfordnlp/dspy
5. TextGrad
Improve prompts and multi-step AI workflows using feedback-inspired optimization.
github.com/zou-group/textgra…
6. EvoAgentX
Build, evaluate, and optimize multi-agent workflows.
github.com/EvoAgentX/EvoAgen…
7. ADAS
Explore how AI can automatically discover and improve agent designs.
github.com/ShengranHu/ADAS
8. OpenEvolve
Use evolutionary search to generate, test, and refine code candidates.
github.com/codelion/openevol…
9. SWE-smith
Generate software-engineering tasks and datasets for training coding agents.
github.com/SWE-bench/SWE-smi…
10. AutoGPT
Experiment with autonomous agents that plan tasks, use tools, and execute multi-step workflows.
github.com/Significant-Gravi…
The interesting shift is this:
AI that writes code → AI that tests ideas → AI that helps discover better solutions.
We’re still early, but these projects offer a glimpse of what comes next.
Bookmark these repos. Which one would you experiment with first?
Latent Node retweeted
Looks like you copied Chandra OCR 2 into your repo, renamed it, changed the license, and made the output worse (your olmocr score is 83.8 vs our 85.8).
Introducing interfaze-1-lite, the first general open-weight model for deterministic work
OCR, speech-to-text, extraction, classification and more
It returns confidence scores and metadata like bounding boxes on every run
huggingface.co/interfaze-ai/…
Readers added context they thought people might want to know
Interfaze appears to have copied the Chandra OCR 2 model, renamed it, and changed the license, without disclosing that the model is not orginal. See x.com/VikParuchuri/s…
Latent Node retweeted
Very good humanizer, I've been trying to train the AI detector against it for several days now and progress is moderate
This matches our experience, frontier models are quite good at detecting AI writing. Including their own writing, they do tend to struggle to fix it or make it more natural or human sounding. That was one of the reason to explore building our own text restyler model - meragpt.com/models/text-rest…
Detecting is easy but making prose human like is still quite hard. Another open model that does well in human-like prose writing is Hemmingway-1 mindstudio.ai/blog/hemmingwa…
Latent Node retweeted
We build Frame for the same find, exploit, fix loop, with the LLM grounded in taint tracking and Z3 proofs. On XBOW/ZeroPath's 39 apps it detects 78.8%, level with ZeroPath and ahead of Semgrep and Snyk, on open models. github.com/lambdasec/frame
Latent Node retweeted
The most striking thing about this is the number of people responding that work in the AI space and seem completely unaware of any type of language model outside of decoders and don’t seem to understand how a model can output probabilities.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Latent Node retweeted
pov: you’re about to read the most confidently incurious, transparently self-serving, afternoon-ruining take you’ve ever seen
Latent Node retweeted
@MovingAtomsLab (a @ycombinator company) has been banned from the Physics-IQ Verified benchmark. The benchmark is a collaboration between @GoogleDeepMind/@Anates_Labs. The ban lasts for 3 months, and their submissions have been rejected. reasons (1/n) 🧵: @bradflora @garrytan
Latent Node retweeted
You Don't need to pay for your next passport photo.
Upload a photo → choose your country → get the right passport dimensions.
Free. No signup. Runs on your device.
And that's just 1 of 32 free tools on MeraSaaS.
→ merasaas.com
Latent Node retweeted
Less setup. More progress. 🚀
Building a startup shouldn’t mean paying for dozens of tools.
MeraSaaS brings the everyday essentials together:
🔐 Authentication
💾 Storage & databases
💳 Payments
Spend more time building.
→ merasaas.com
Latent Node retweeted
OpenEvolve turns LLMs into autonomous code optimizers that discover new algorithms.
github.com/algorithmicsuperi…
Latent Node retweeted
My Apple Silicon runtime search is finally over!✅
Tested MLX, llama.cpp, Ollama, and LM Studio extensively.
Winner: Qwen3.6-35B-A3B-OptiQ-4bit + oMLX 🏆
M3 Ultra 96GB:
⚡1145.5 tok/s prompt
🔥68 tok/s generation
📚65.5K context (Hermes Agent)
Big thanks to @jundotkim🙏
Latent Node retweeted
The "Qwen3.6-35B-A3B-OptiQ-4bit" is the most reliable local model with reasonable performance all round
Latent Node retweeted
the mlx 4bit quant i used mlx-optiq.com/docs/qwen3.6 which follows a different dynamic quantization from unsloth: unsloth infers importance from activations + finer block control using a 1.5M calib tokens; optiq measures it via 40 curated samples and directly + solves a BPW budget in round-trip, whole-layer flips (4 or 8), no block mixing.
Latent Node retweeted
Quantize LLMs per layer, not per model.
mlx-optiq is mixed-precision quantization for Apple Silicon:
- Per-layer precision tuning
- Built for MLX, runs on M-series chips
- Drops model size without the quality cliff
The Apple Silicon quantization tool.
mlx-optiq.com/
Latent Node retweeted
Interesting architectural finding from running OptiQ on LFM2.5-8B-A1B (18 Conv + 6 GQA layers):
The KL-sensitivity sweep consistently pushes every MoE feed-forward gate to 8-bit precision.
Why?
The gating layers determine which experts are activated for each token. Even small quantization errors in these layers can alter routing decisions, and those routing changes compound throughout generation. The result is a disproportionately large impact on model quality relative to their size.
OptiQ's solution is straightforward:
• Keep the MoE gates at high precision (8-bit)
• Aggressively compress the expert weights instead
In other words, spend bits where routing decisions are made and save bits where the bulk of the parameters live.
The model card includes:
• Quality benchmark results
• Full per-layer bit allocation table
• Quantization analysis
huggingface.co/sahilchachra/…
