KABLAMO · APPLIED AI
On Our Minds
AI moves fast. These are the announcements our engineers have been reading, sharing and testing over the last 30 days: new models, skill repositories, and posts from partners and peers.
thinkrail: Vibe code with pi in a lightweight, real IDE that customises itself around the way you work
JetBrains' open-source, lightweight IDE for vibe coding with pi, which adapts itself to the way you work.
OpenAI and Anthropic wanted to replace engineers. Instead, they shot their own customers.
An essay arguing that cheap AI coding lets anyone clone expensive software, which undermines the businesses AI labs hoped would pay them.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model
Moonshot AI says its new open-weight MoE model has a 1M-token context and handles full-project coding, 3D modelling and finance work.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index
Artificial Analysis says Kimi K3.1 Pro scores 55 on its index, 11 points above Kimi K3, with large gains in agentic performance.
We're adding a new plugin to Claude Code: You should Know
Anthropic says the built-in plugin scans Claude's output for important information you might miss and is enabled with one /plugin command.
How Claude remembers your project
Covers CLAUDE.md, AGENTS.md and auto memory, plus /doctor prompt-audit, which flags outdated, missing or conflicting instructions and proposes edits.
Markdown is the new executable: attackers are dropping agent instructions into your Downloads folder from the web
Forever found live malware that plants a file in the Downloads folder to turn a developer's own coding agent against them, and shows how it works on Codex and Claude Code.
We're opening a research preview of Claude Duet, a new app where Claude works as a proactive agent
Anthropic says Claude Duet brings you what needs your attention, does tasks you ask for, keeps working while you're away and checks with you before acting.
How to run Turing-complete malware in PyTorch's backward pass
Elon Litman uses a floating-point exploit to turn PyTorch's backward pass into a Turing-complete virtual machine.
DeepSeek V4.1 Flash 1017 runs at 1,521+ tok/s on Fireworks Ultrafast
Fireworks says its Ultrafast tier runs DeepSeek V4.1 Flash at over 1,521 tokens per second, and it has opened a waitlist.
Agentic AI coding breaks mob programming (here's how to fix it)
Technical coach Emily Bache looks at how agentic AI coding disrupts mob and pair programming, and how teams can adapt.
We think that as early as 2027 or 2028, there might be a major AI-enabled terrorism event
Stephen Casper shares a new paper: a pre-mortem on AI-enabled terrorism, with nine falsifiable predictions and two concrete scenarios.
Steering Claude Code: when to use CLAUDE.md, skills, hooks, and subagents
Seven ways to steer Claude Code, including CLAUDE.md files, rules, skills, subagents and hooks, and when to use each based on context cost and authority.
Announcing AWS Well-Architected Agent, an AI-powered intelligence to optimize your cloud environment (preview)
A public preview of an AI service that analyses your AWS environment and gives targeted recommendations for improving your applications.
Adding Pyrefly type checking to your agentic loop
How to use skills and hooks to add Pyrefly type checking to AI agent workflows, so generated Python code is checked automatically.
Notice of order regarding exclusion and removal of DeepSeek products from federal civilian executive branch information technology systems and procurements
A DHS order that bars DeepSeek products from US federal civilian procurement and removes them, and third-party products built on them, from agency IT.
I built a programming language that LLMs can't use
KillSwitch is an esoteric language designed to confound LLMs and coding agents, with benchmark results for the major models against it.
We placed the fruit fly brain inside Odyssey-3, a frontier world model
Odyssey says it placed a fruit fly brain inside its Odyssey-3 world model, where it now flies around simulated worlds by itself.
We had a security breach last night related to a large scale distillation attack
OpenCode says a large-scale distillation attack affected 1.4% of OpenCode Console users, all with external providers linked, and that it has emailed them.
Harness^2: Recursive agent harnessing for an open world
A framework that improves an agent's harness while it works, without ground-truth checks, reporting gains of up to 11.8 points across three professional benchmarks.
GLM 5.3 Flash is now free in OpenCode
OpenCode says GLM 5.3 Flash is now free to use in its coding agent.
Today we're releasing Muse Spark Open, the most capable US open-weight model to date
Meta says Muse Spark Open, an open-weight model with 816B parameters and a 1M-token context window, is built for coding and agentic work.
NVIDIA releases 2 new models: Nemotron-3.5-Bolt and Nemotron-3.5-Super-VL
Unsloth AI says Nvidia's Nemotron-3.5-Bolt runs in 24GB of RAM and Super-VL in 82GB, and has published GGUF builds and a guide.
open-code-review: Secure, fast, efficient, battle-tested at Alibaba's scale
Alibaba's code review tool, which pairs fixed review steps with an LLM agent to leave line-level comments, using built-in rules for common bugs and vulnerabilities.
Human–AI collaboration at scale: task criticality, agency, and friction across 250,000 conversations
Stanford researchers studied nearly 250,000 real Claude.ai conversations to see how people and AI work together.
OpenCode 2.0
Release notes for OpenCode 2.0.
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work
Daanish Khazi says the open-source Ultramarine 1 Flash matches Fable 5 at a sixtieth of the cost, with weights due in two weeks.
Here is Qwen 3.8 27B compared to all the SOTA models 1 year ago today on Artificial Analysis
Youssof Al Toukhi says Qwen 3.8 27B would have been the best model in the world a year ago.
New on our Engineering blog: Creating in-house source control
Anthropic explains why it built its own source control after agents pushed merges from about 2,500 to more than 30,000 changes a day.
Gemini 4 Argon is now GA
Artificial Analysis says Gemini 4 Argon ranks level with GPT-6 Astra on its Intelligence Index at 60% of the cost per task.
Gemini 4 Argon: built to solve your hardest problems
Google's latest frontier model, now available to developers, enterprises, scientists and consumers.
AI models may be capable of self-replicating without access to a datacenter
Anthropic research suggesting small models may be months away from self-replicating across cheap IoT devices such as Android TV boxes.
Deep dive into Azure Maia 300 architecture
A look inside Microsoft's Maia 300 AI chip and how it aims to make advanced AI capabilities more accessible and affordable.
Intel accelerates open infrastructure for the agentic AI era
Intel's GPU D-series, Xeon CPUs, open software and ecosystem updates aimed at helping customers scale agentic AI.
Deploy Langfuse on GCP with Terraform
A step-by-step guide to self-hosting Langfuse on Google Cloud with Terraform.
Why OpenAI's latest results left mathematicians in shock
An Information Age piece asking whether OpenAI's recent maths results hold up under expert scrutiny.
Coming today! Beat DGX Spark! MaixStation soon
Sipeed teases its upcoming MaixStation hardware, claiming it will beat DGX Spark.
Introducing Muse Wonder and Muse Wonder Full: frontier models for agents
Meta's most capable models to date, built for state-of-the-art performance on personal and professional agentic tasks.
Meta has released Muse Wonder and Muse Wonder Full
Artificial Analysis scores Muse Wonder and Muse Wonder Full at 52 and 56 on its Intelligence Index, up from 48 for Muse Spark 1.3.
Fine-tuning an open-weight model (Qwen3.5-9B) on proprietary internal data at Coinbase
Brian Armstrong says a Qwen3.5-9B model fine-tuned on Coinbase's internal data outperformed frontier models at fraud prevention, at much lower cost.
Introducing Multi-Agent Self-Supervision (MASS): recursive self-improvement through collective intelligence
Sakana AI introduces MASS, a recursive self-improvement method for open-ended research problems where human experts may struggle to assess results.
Pensieve: local code search and MCP context for AI coding agents
A free, local tool that searches everything on your Mac by meaning, giving coding agents context in one search instead of a disk-wide grep.
rea: reverse engineer anything with agents, from app behavior down to native binaries
An open-source repo for using agents to reverse engineer software, from how an app behaves down to its native binaries.
jev-ultrafast: the fastest and cheapest web agent
Browser Use's web agent, built to be the fastest and cheapest way to get work done in a browser.
Model-based grading
Asking an LLM judge for strengths and weaknesses alongside the score stops it defaulting to middling marks.
Dark Testing Factory: a self-operating quality system
UiPath's white paper on a quality system that plans, designs, runs and maintains software tests end to end without people in the loop.
Pi 1.0
A minimal, extensible agent harness hits 1.0, alongside Pi Durable, an experimental base for long-running agent applications.
DeepSeek Harness desktop app
DeepSeek's harness gets a desktop app for working with documents, spreadsheets and code, extended through plugins.
Comfy Agent
An agent that builds, runs and reviews ComfyUI workflows while you edit the canvas. Live in Comfy Cloud, with local support coming.
Advent of Agents, season 3
31 days of hands-on tutorials this October for building production-ready agents with Gemini, ADK and Vertex AI.
Engram
DeepSeek's paper on conditional memory through scalable lookup: a new way to make large language models sparse.
Anthropic publishes prompting guidance for Claude Opus 5.5
Anthropic's guide for Opus 5.5 asks developers to retest effort settings and rethink prompts that tell Claude to think carefully.
GLM-5.3 and the spread of advanced cyber capabilities
GLM-5.3 can build end-to-end cyber exploits on its own, and unlike other frontier models it shipped without meaningful safeguards.
OpenAI business marketplace
OpenAI's marketplace for business.
Lookalike letters don't fool GPT-6 or Claude, but nearly quadruple the bill
Lookalike Unicode letters don't trick current GPT-6 or Claude models, but reading them costs up to 5.7x the tokens and 3.9x the bill.