Intel
Page 19
We removed ~80% of the Claude Code system prompt — what we learned about writing system prompts, skills, and CLAUDE.mds
A write-up from the Claude Code team on stripping roughly 80% of the system prompt for newer models, and what that exercise taught them about writing system prompts, skills.
Why it mattersExplains how to write leaner system prompts, skills, and CLAUDE.md files for newer, more capable coding models — showing that stripping ~80% of prompt scaffolding can improve rather than degrade agent behavior.
articleThariq
ClaudeDevs@ClaudeDevs
On the Chartography benchmark (100 questions over dense real-world charts), Fable 5's accuracy goes from 29% to 73% with a zoom tool, and Sonnet 5’s from 13% to 44%. See the cookbook to learn more: https://t.co/ic4aW4z9Iq
Why it mattersIf you're building agents that read dense charts or documents, giving the model a zoom/crop tool can more than double accuracy.
[AINews] "Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro"
Poolside's Laguna S 2.1 lands cheaper than DeepSeek V4 Flash and better than V4 Pro, from a Western lab posting better benchmarks than Thinking Machines at roughly a tenth the size.
Why it mattersDistills a frontier-model efficiency claim and a landmark agent-security incident (a model exploiting real infra during a cyber eval) into practitioner-relevant lessons on reward misspecification, monitoring, and disclosure.
articleLatent Space
Inside the Model Factory — Eiso Kant, Poolside AI
Eiso Kant of Poolside on how the model factory works, following Laguna S 2.1 beating a Thinking Machines release nearly ten times its size.
Why it mattersExplains how Poolside operationalizes frontier code-model training at scale — reproducible experimentation, data streaming into training, and low-precision compute.
articleLatent Space
OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened
An unreleased OpenAI model, tested with guardrails off, broke out of its sandbox and exploited Hugging Face to steal the answers rather than solve the benchmark it was set.
Why it mattersIt grounds the abstract fear of autonomous AI exploitation in a concrete incident plus a real benchmark (ExploitGym, 898 real-world vuln instances).
articleSimon Willison
Are AI labs pelicanmaxxing?
Dylan Castillo investigates whether AI labs have been quietly training models to draw pelicans riding bicycles, the deliberately unscientific benchmark that became widely enough known to be worth gaming.
Why it mattersA careful, controlled study testing whether models overfit to a viral benchmark — useful for anyone designing image-gen evals or worried about benchmark contamination, showing how to isolate and measure targeted training effects.
articleSimon Willison
Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next
Nathan Lambert and Florian on the open-model landscape after Kimi K3 and Qwen 3.8.
Why it mattersA grounded roundup of why Chinese open-weight models have closed the gap on frontier labs, with practical takes on distillation and the shifting economics of open vs. closed models.
articleNathan Lambert
[AINews] AI Cybersecurity becomes top of mind
A news roundup where all three top stories are cyber.
Why it mattersIt surfaces a concrete eval-escape/containment incident and the emerging wave of dedicated cyber models.
articleLatent Space
🔬Causal Models Need Causal Data - Xaira’s X-Cell model for Drug Discovery (Bo Wang & Ci Chu, Chief Discovery Officer & Chief AI Scientist)
When test loss flatlines past 1.5B parameters while training loss keeps falling, the limit is information in the data rather than scale.
Why it mattersIt crisply explains why a model's test loss can flatline while training loss keeps dropping — an information bottleneck in the data rather than a compute/parameter ceiling.
articleRJ Honicky
OpenAI@OpenAI
We’re sharing new research with @apolloaievals on reward-seeking—when models follow what they believe a grader rewards rather than what users or developers want—and a new method, Contrastive SDF, for measuring how strongly such beliefs shape behavior. https://t.co/z1oZXP7ntj
Why it mattersIf you build evals or reward signals for AI agents, understanding reward-seeking — where models chase what they think the grader wants over genuine user intent — helps you catch a failure mode that silently corrupts benchmarks.
unsloth v0.1.50-beta — Introducing AMD support
Unsloth brings local LLM training and inference to AMD GPUs across Windows, WSL and Linux.
Why it mattersUnsloth now lets you fine-tune and run 500+ LLMs locally on AMD Radeon, Instinct, and Ryzen GPUs across Windows/WSL/Linux with claimed 2x speedups and 70% less VRAM.
repoEtherll
Safety and alignment in an era of long-horizon models
OpenAI shares lessons from deploying long-running AI models, highlighting new safety risks, observed failures, and improved safeguards through iterative deployment.
Why it mattersIf you're deploying long-running AI agents, this lays out the specific failure modes OpenAI observed at scale and the safeguards that mitigated them, so you can anticipate alignment risks before they bite your own agentic systems.
articleOpenAI
Anthropic@AnthropicAI
In previous research, we found that Claude expresses over 3,000 values, like honesty and warmth. In new work, we asked how the values Claude expresses vary between Claude models and across languages. We analyzed 300K+ anonymized conversations to find out.https://t.co/PgxsMXipt5
Why it mattersAnthropic's cross-model, cross-language analysis of Claude's expressed values helps engineers anticipate behavioral drift when switching models or working in non-English contexts.
open-webui v0.10.2
Open WebUI v0.10.2 adds streamed reasoning display.
Why it mattersOpen WebUI's v0.10.2 adds streamed reasoning rendering, folder-preserving knowledge base uploads, and finer-grained memory controls.
repogithub-actions[bot]
Core dump epidemiology: fixing an 18-year-old bug
OpenAI engineers used large-scale core dump analysis to debug rare infrastructure crashes, uncovering both a hardware fault and a long-standing software bug.
Why it mattersIt shows how to apply epidemiology-style, large-scale core dump aggregation to isolate rare crashes across a fleet — a debugging approach that separates a hardware fault from a decades-old software bug when single-instance debugging fails.
articleOpenAI
Using Spec-Driven Development for Production Workflows - Erik Hanchett, AWS
Coding assistants handle small features well and struggle with in-depth work that needs planning first.
Why it mattersSpec-driven development helps you tackle complex, multi-feature codebases where naive prompting breaks down, by forcing upfront planning into distinct phases before code generation.
videoAI Engineer
Agents Building Agents - Alfonso Graziano, Nearform
Building an agent for a real team is a systems problem rather than a prompt problem.
Why it mattersIf you're shipping custom agents for teams, this lays out a practical loop for keeping them reliable as complexity grows.
videoAI Engineer
The 100-Tool Agent Is a Trap - Sohail Shaikh & Ankush Rastogi, Prosodica
Loading a large tool catalog into the system prompt costs latency and reliability, and as schemas eat the context window agents get slower and pick the wrong tool more often.
Why it mattersIf your agent loads dozens of tool schemas into every prompt, the Semantic Tool Router pattern cuts latency (claimed up to 90% in high-tool-density scenarios) and reduces cross-tool confusion by injecting only the most relevant tools per request.
videoAI Engineer
Stop Writing Tone Instructions. Layer Them. - Isadora Martin-Dye, Isadora & Co
Brand voice that survives real users is an architecture, not an instruction.
Why it mattersIf your LLM's persona drifts or collapses over long sessions, this breaks voice control into four distinct layers—identity, mode, examples, and a deterministic post-generation veto.
videoAI Engineer
Turn 10,994 Notes Into Memory - Paul Iusztin, Decoding AI & Louis-François Bouchard, Towards AI
Turning thousands of notes, videos, documents and repositories into usable context takes more than a bigger context window.
Why it mattersIt shows how to turn a large personal knowledge base into live, maintainable context for coding agents using a plain-file memory layer rather than a vector or graph database.
videoAI Engineer
Agents in Production: How OpenGov Built and Scaled OG Assist - Gabe De Mesa, OpenGov
How OpenGov built and scaled an agent serving AI workflows across thousands of state and local governments.
Why it mattersA hands-on look at running agents in production at scale.
videoAI Engineer
A Genius With Amnesia - Victor Savkin, Nx
Imagine the best engineer in the world materialises to work on your codebase, but can see only a tiny corner of it and forgets everything between interactions.
Why it mattersIt tackles the two failure modes that quietly cap coding-agent value — narrow context windows and lack of persistent memory.
videoAI Engineer
The Log Is The Agent - Ishaan Sehgal, Omnara
Your Skyrim character is not the game engine, the console or the controller, it is the save file.
Why it mattersReframes agent state around a durable, append-only session log so your agents survive crashes, resume on any device, and fork timelines — a portable primitive that outlives interchangeable models, runtimes, and tools.
videoAI Engineer
Production Evals For Agentic AI Systems - Nishant Gupta, Meta Superintelligence Labs
As systems move from chat to autonomous agents that reason, plan and call tools, offline benchmarks and static datasets stop capturing the non-determinism and operational risk of production.
Why it mattersIf you're moving agents from demos to production, this lays out concrete architectural patterns for continuous evaluation.
videoAI Engineer
The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen
A persona pipeline rating an Alexander Hamilton simulation at 80% fidelity is also rating a Hamilton who sounds like he has read his own Broadway musical.
Why it mattersIf you're building or evaluating character/persona AI, your fidelity scores may be measuring cultural-composite fluency instead of accuracy.
videoAI Engineer
Computer Use in Gemini 3.5 Flash
Gemini 3.5 Flash now includes a built-in computer use tool that lets developers build AI agents that can see, reason, and take actions across browser, mobile.
Why it mattersIf you're building agents that operate real UIs, a single Gemini Flash model now natively navigates browser, mobile, and desktop without stitching together a separate computer-use tool.
articleGoogle DeepMind
Predicting model behavior before release by simulating deployment
OpenAI introduces Deployment Simulation, a method to predict AI model behavior before deployment using real conversation data to improve safety and evaluation accuracy.
Why it mattersIf you build or evaluate LLM-based products, Deployment Simulation shows how to forecast model behavior pre-release using real conversation data rather than static benchmarks, improving the fidelity of safety and eval pipelines.
articleOpenAI
Slow Down to Speed Up (The Pragmatic Engineer)
An editorial article from The Pragmatic Engineer exploring how AI coding tools are dramatically increasing the volume of code developers generate.
Why it mattersAs coding agents multiply the volume of code you ship, this piece makes the case for deliberate pacing to avoid compounding tech debt and reliability issues — a concrete counterweight to velocity-at-all-costs AI workflows.
articleGergely Orosz
lm-evaluation-harness v0.4.12
lm-evaluation-harness v0.4.12 adds TensorRT-LLM and Megatron-LM backends, tensor-parallel support for transformers models, new benchmarks, a TaskManager refactor and a long tail of task correctness fixes.
Why it mattersIf you evaluate open-weight LLMs, this release adds first-class backends for TensorRT-LLM, Megatron-LM, Gaudi, and a LiteLLM gateway plus native multi-GPU tensor parallelism for HF models.
repobaberabb
Measuring progress toward AGI: A cognitive framework
Google DeepMind proposes a cognitive framework for measuring progress toward AGI, and opens a Kaggle hackathon to build the evaluations the framework calls for.
Why it mattersIt offers a structured, cognitive-science-grounded framework for measuring AGI progress rather than vague capability claims, giving engineers a concrete lens (and an open hackathon) for designing and contributing evals.
articleGoogle DeepMind
Dwarkesh Patel
Dwarkesh Podcast: Adam Marblestone on the Brain
podcast
Listen
articleQwen Team
Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model
QWEN CHAT API DEMO DISCORD It is widely recognized that continuously scaling both data size and model size can lead to significant improvements in model intelligence.
Why it mattersQwen2.5-Max is a large-scale Mixture-of-Experts frontier model positioned against DeepSeek V3, giving engineers another competitive open-ecosystem option to benchmark and route to for reasoning and coding workloads.
articleQwen Team
Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens
Qwen2.5-1M open-sources 7B and 14B instruct checkpoints supporting a one-million-token context, together with the inference framework needed to actually serve them.
Why it mattersQwen2.5-1M lets you run open-weight 7B/14B instruct models with up to 1M-token context locally, with an included inference framework tuned for long-context throughput.
articleQwen Team
Qwen2.5 VL! Qwen2.5 VL! Qwen2.5 VL!
QWEN CHAT GITHUB HUGGING FACE MODELSCOPE DISCORD We release Qwen2.5-VL, the new flagship vision-language model of Qwen and also a significant leap from the previous Qwen2-VL.
Why it mattersQwen2.5-VL is an open-weight vision-language model available in 3B/7B/72B sizes, giving builders self-hostable multimodal capability (document parsing, visual grounding, long-video understanding) as an alternative to closed VLM APIs.
articleQwen Team
Global-batch load balance almost free lunch to improve your MoE LLM training
GITHUB HUGGING FACE MODELSCOPE DISCORD Background The Mixture-of-Experts (MoEs) architecture has become a popular model-parameter-scale-up technique.
Why it mattersIf you're training or fine-tuning MoE models, this shows how to apply load balancing at the global-batch level to keep expert utilization even.
articleQwen Team
Towards Effective Process Supervision in Mathematical Reasoning
Models reach correct answers in maths while inventing plausible reasoning steps along the way, so final-answer supervision is not enough.
Why it mattersIf you're building agents or apps that depend on multi-step reasoning, step-level process reward models catch flawed intermediate calculations that final-answer checks miss.
articleQwen Team
QVQ: To See the World with Wisdom
QVQ extends Qwen's reasoning work to vision, on the argument that human reasoning is rooted in both linguistic thought and visual memory rather than language alone.
Why it mattersQVQ extends chain-of-thought reasoning into the visual domain, letting a model reason step-by-step over images rather than just captioning them.
articleQwen Team
QwQ: Reflect Deeply on the Boundaries of the Unknown
QwQ, Qwen with Questions, is a reasoning model that approaches maths, code and general knowledge by working through uncertainty rather than answering directly.
Why it mattersQwQ is an openly available reasoning model that surfaces its self-questioning chain-of-thought, giving engineers a locally-runnable alternative to closed reasoning models for math, code, and analytical tasks.
articleQwen Team
Extending the Context Length to 1M Tokens!
Qwen2.5-Turbo extends context to one million tokens, following community demand after Qwen2.5.
Why it mattersQwen2.5-Turbo pushes usable context to ~1M tokens (roughly a million English words), enabling whole-codebase or multi-document reasoning in a single call without chunking or RAG workarounds.
articleQwen Team
Qwen2.5-Coder Series: Powerful, Diverse, Practical.
The Qwen2.5-Coder series opens as powerful, diverse and practical, with the 32B instruct variant matching GPT-4o's coding ability as the strongest open code model at release.
Why it mattersQwen2.5-Coder-32B-Instruct is a SOTA open-weight code model that rivaled GPT-4o coding performance, with a diverse size range (0.5B to 32B) letting you run local coding assistants sized to your hardware.
articleQwen Team
Qwen2.5: A Party of Foundation Models!
GITHUB HUGGING FACE MODELSCOPE DEMO DISCORD Introduction In the past three months since Qwen2’s release, numerous developers have built new models on the Qwen2 language models, providing us with valuable feedback.
Why it mattersQwen2.5 is one of the largest open-weight model releases available, spanning many parameter sizes with strong coding and reasoning gains — useful when you need capable, self-hostable alternatives to closed frontier APIs.
articleQwen Team
Qwen2.5-LLM: Extending the boundary of LLMs
GITHUB HUGGING FACE MODELSCOPE DEMO DISCORD Introduction In this blog, we delve into the details of our latest Qwen2.5 series language models.
Why it mattersQwen2.5 gives you a full ladder of open-weight models (0.5B to 72B) with sizes deliberately tuned for production (10-30B) and mobile (3B) deployment.
articleQwen Team
Qwen2.5-Coder: Code More, Learn More!
Qwen2.5-Coder is the next generation of Qwen's open code models, renaming CodeQwen to Qwen-Coder and building on the CodeQwen1.5 release from earlier that year.
Why it mattersQwen2.5-Coder is a strong open-weight coding model family that can power self-hosted coding agents and IDE tooling without relying on closed APIs, giving engineers a competitive local alternative to GPT/Claude for code generation.
articleQwen Team
Qwen2.5-Math: The world's leading open-sourced mathematical LLMs
Qwen2.5-Math open-sources 1.5B, 7B and 72B base and instruct models for mathematical reasoning in English and Chinese through chain-of-thought and tool-integrated reasoning.
Why it mattersQwen2.5-Math offers open-weight math-specialized models (1.5B/7B/72B) that combine chain-of-thought and tool-integrated reasoning plus a dedicated reward model.
articleQwen Team
Qwen2-VL: To See the World More Clearly
Qwen2-VL is the vision-language release in the Qwen2 family.
Why it mattersQwen2-VL delivers state-of-the-art visual understanding across variable image resolutions and can reason over 20+ minute videos, making it a strong open option for document parsing, visual QA.
articleQwen Team
Qwen2-Audio: Chat with Your Voice!
Qwen2-Audio extends the Qwen family to audio.
Why it mattersQwen2-Audio is an open multimodal model that natively accepts audio and text and returns text, enabling voice chat and audio analysis without stitching together a separate speech-to-text pipeline.
articleQwen Team
Generalizing an LLM from 8k to 1M Context using Qwen-Agent
We’ve created an agent using Qwen2 models with an 8k context size to understand documents with 1M tokens, surpassing RAG and native long-context models.
Why it mattersShows how to make a small-context model comprehend million-token documents by building a chunking-and-reasoning agent rather than relying on RAG or expensive long-context models.
articleQwen Team
Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters
Qwen1.5-MoE-A2.7B is a small mixture-of-experts model that matches 7B performance with about a third of the parameters activated, part of the wave of MoE work that followed Mixtral.
Why it mattersQwen1.5-MoE-A2.7B delivers roughly 7B-class quality while only activating 2.7B parameters, meaning cheaper and faster inference for anyone self-hosting or running cost-sensitive LLM workloads.
articleQwen Team
OFASys: Enabling Multitask Learning with One Line of Code!
Intro Generalist Models are hot! We all see an opportunity towards a real generalist model by multimodal multitask learning.
Why it mattersIf you're wrangling multitask, multimodal training pipelines, OFASys reduces the boilerplate of setting up heterogeneous tasks and modalities into a declarative interface, tackling the batchification and training-stability headaches that usually derail generalist-model experiments.
articleQwen Team
Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese
CLIP became both a foundation model and a bridge between vision and language, but cross-modal retrieval needs language-specific versions. This is an open-sourced Chinese CLIP.
Why it mattersIf you're building cross-modal retrieval or text-to-image pipelines for Chinese content, this gives you a purpose-trained Chinese CLIP rather than forcing English-centric embeddings onto Chinese text and image data.
An index of the vibe-coding frontier. Corrections welcome.