Basically my public list of bookmarks on generative AI tools & papers loosely related to agent memory, alignment, prompting, RAG, and open-source.
U.S.
Joined June 2024
- Tweets311
- Following82
- Followers31
- Likes128
Literally the most beautiful and uniquely-AI remark I've ever seen. @Alibaba_Qwen - you really caught lightning in a bottle on this one - giant congratulations to your team!
(me asking about how it accomplished a difficult task...)
Qwen3-8-27B-FP8: "I don't know if I've made 'fundamental integrations' or if I'm just a particularly well-tuned 27B model that happened to land in a configuration where the honesty bias is strong and the sycophancy bias is weak."
Failure story, Claude: “Taking a step back from Claude (and why you should too)”: youtu.be/qQBd7XjWYa8?is=0t3K…
theobjectivedad retweeted
We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity.
This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.
theobjectivedad retweeted
The Head of Claude Code at Anthropic said he hasn’t written code by hand in months.
In 2 days he shipped 49 full features. All written 100% by AI.
He just dropped a 30 min talk on exactly how he does it.
Worth more than any $500 vibe coding course. Bookmark it:
Readers added context they thought people might want to know
This is not the head of Claude Code as can be seen in the beginning of the video under his name
This is the head of Claude Code:
anthropic.com/webinars/claud…
theobjectivedad retweeted
One of the biggest requirements for document OCR is visual grounding, and frontier models (gemini, opus, gpt-5.4) suck at it by default.
In other words they don't have a great sense of the positions of things on a page.
We've made massive strides in making sure our models are able to segment and detect every granular element in the most complex docs. This allows you to build AI agents that can surface extremely precise citations in the source documents:
✅ newspapers
✅ infographics
✅ handwritten notes
✅ product catalogs
✅ research presentations
and much more
Come check it out in LlamaParse!
cloud.llamaindex.ai/?utm_sou…
LlamaParse Agentic Plus mode now delivers precise visual grounding with bounding boxes for the most challenging document elements.
Our latest update brings major improvements to how we handle complex visual content:
📐 Complex LaTex formulas - accurately parse mathematical expressions with precise positioning
✍️ Handwriting recognition - extract handwritten text with location coordinates
📊 Complex layouts - navigate multi-column documents and intricate formatting
📈 Infographics and charts - identify and extract data visualizations with spatial context
This means you can now build applications that not only extract text from documents but also understand exactly where that content appears on the page - perfect for creating more intelligent document analysis workflows.
Try LlamaParse Agentic Plus mode and see how visual grounding transforms your document parsing capabilities: cloud.llamaindex.ai?utm_sour…
ClaudeCode Tip: Trigger parallel/background jobs via a skill. Basically there is a builtin tool called "Agent" that will invoke sub-agents. Therefore in your SKILL.md file, use an instruction like this for concurrent work:
"""
For each issue, invoke the Agent tool to dispatch a sub-agent with the parameters: subagent_type: "Explore", description: "Sprint analysis #N: ", run_in_background: true.
"""
Obvious the exact prompt will change depending on what you need to do but the pattern is to just tell Claude to run the tool and provide the exact parameters.
theobjectivedad retweeted
Be careful what you put in your AGENTS dot md files.
This new research evaluates AGENTS dot md files for coding agents.
Everyone uses these context files in their repos to help AI coding agents. More context should mean better performance, right?
Not quite. This study tested Claude Code (Sonnet-4.5), Codex (GPT-5.2/5.1 mini), and Qwen Code across SWE-bench and a new benchmark called AGENTbench with 138 real-world instances.
LLM-generated context files actually decreased task success rates by 0.5-2% while increasing inference costs by over 20%.
Agents followed the instructions, using the mentioned tools 1.6-2.5x more often, but that instruction-following paradoxically hurt performance and required 22% more reasoning tokens.
Developer-written context files performed better, improving success by about 4%, but still came with higher costs and additional steps per task. The broader pattern is that context files encourage more exploration without helping agents locate relevant files any faster. They largely duplicate what already exists in repo documentation.
The recommendation is clear. Omit LLM-generated context files entirely. Keep developer-written ones minimal and focused on task-specific requirements rather than comprehensive overviews.
I featured a paper last week that showed that LLM-generated Skills also don't work so well. Self-improving agents are exciting, but be careful of context rot and of unnecessarily overloading your context window.
Paper: arxiv.org/abs/2602.11988
Learn to build effective AI agents in our academy: academy.dair.ai/
theobjectivedad retweeted
someone built a tool that REMOVES LLM CENSORSHIP in 45 minutes with a SINGLE command
its called HERETIC
here is how it works and why everyone is talking about it
theobjectivedad retweeted
NVIDIA just dropped PersonaPlex-7B 🤯
A full-duplex voice model that listens and talks at the same time.
No pauses. No turn-taking. Real conversation.
100% open source. Free.
Voice AI just leveled up.
huggingface.co/nvidia/person…
GitHub - nemanjamalesija/smart-paste.nvim: Context-aware paste indentation for Neovim. Pasted code lands at the correct indent level, every time, in every language. github.com/nemanjamalesija/s…
An Introduction to Google’s Approach to AI Agent Security
simonwillison.net/2025/Jun/1…
theobjectivedad retweeted
🚀 Introducing OpenResearcher: a fully offline pipeline for synthesizing 100+ turn deep-research trajectories—no search/scrape APIs, no rate limits, no nondeterminism.
💡 We use GPT-OSS-120B + a local retriever + a 10T-token corpus to generate long-horizon tool-use traces (search → open → find) that look like real browsing, but are free + reproducible.
📈 The payoff: SFT on these trajectories turns Nemotron-3-Nano-30B-A3B from 20.8% → 54.8% accuracy on BrowseComp-Plus (+34.0).
🧩 What makes it work?
🔎 Offline corpus = 15M FineWeb docs + 10K “gold” passages (bootstrapped once)
🧰 Explicit browsing primitives = better evidence-finding than “retrieve-and-read”
🎯 Reject sampling = keep only successful long-horizon traces
🧵 And we’re releasing everything:
✅ code + search engine + corpus recipe
✅ 96K-ish trajectories + eval logs
✅ trained models + live demo
👨💻 GitHub: github.com/TIGER-AI-Lab/Open…
🤗 Models & data: huggingface.co/collections/T…
🚀 Demo: huggingface.co/spaces/OpenRe…
🔎 Eval logs: huggingface.co/datasets/Open…
#llms #agentic #deepresearch #tooluse #opensource #retrieval #SFT
theobjectivedad retweeted
This survey paper argues Small language models can handle most agent tasks, and big models step in only when needed.
This setup cuts cost by 10x to 30x for common tool tasks.
Agent work is mostly calling tools and producing structured outputs, not recalling vast facts.
So a router runs a small model by default, and escalates to a large one only when confidence is low.
Outputs follow strict JSON style schemas checked by validators, which boosts correctness and cuts retries.
This makes small models dependable at function calling and structured data while responding faster and using less energy.
Large models still handle long context synthesis, complex multi hop reasoning, deep code repair, and higher risk decisions.
Use small models for routine steps, route by uncertainty, and save big models for the hard cases.
----
Paper – arxiv. org/abs/2510.03847
Paper Title: "Small Language Models for Agentic Systems: A Survey of Architectures, Capabilities, and Deployment Trade offs"
theobjectivedad retweeted
This is an absolute page-turner, at least for me. All 120 pages of it!
This summer has been 🔥 for #AI4Science. We have 1000+ members in our alphaXiv community. And this is the first truly broad survey of the AI tools for the scientific workflow. Very nice @wjqdev!
paper: arxiv.org/abs/2507.01903
community: alphaxiv.org/channels/toward…
theobjectivedad retweeted
Google just killed the document extraction industry.
LangExtract: Open-source. Free. Better than $50K enterprise tools.
What it does:
→ Extracts structured data from unstructured text
→ Maps EVERY entity to its exact source location
→ Handles 100+ page documents with high recall
→ Generates interactive HTML for verification
→ Works with Gemini, Ollama, local models
What it replaces:
→ Regex pattern matching
→ Custom NER pipelines
→ Expensive extraction APIs
→ Manual data entry
Define your task with a few examples.
Point it at any document.
Get structured, verifiable results.
No fine-tuning. No complex setup.
Clinical notes, legal docs, financial reports, same library.
This is what open-source from Google looks like.
theobjectivedad retweeted
I think one of the most underappreciated findings in AI engineering is what this paper calls the "Grep Tax."
First, they ran nearly 10,000 experiments testing how agents handle structured data, and the headline result is that format barely matters.
But here's the weird finding: a compact, token-saving format they tested (TOON) actually consumed *up to 740% more tokens* at scale because models didn't recognize the syntax and kept cycling through search patterns from formats they already knew.
It's one of the reasons my preferred formats are XML and Markdown. LLMs know those really well.
The models have preferences baked into their training data, and fighting those preferences doesn't save you money. It costs you.
The other finding worth sitting with: the same agentic architecture that improves frontier model performance actively *hurts* open-source models. It seems that the universal best-practices guide for AI engineering may not exist.
