My feed has been overrun by mentions of Jev from TypeSafe AI - Introducing System One Models and Jev and many folks are stuffing it into every use case imaginable while others are claiming that it is “just” another classifier. I did a little bit of experimentation over the weekend: What it IS: Jev is closest to what is called a Zero Shot Classifier that is hosted with a low latency API around it (having a classical ML background helps here). This is not trivial - Classifiers have been around for a long time but a general purpose Classifier => Any classifier task zero shot => No need to provide examples first hosted low latency=> Fast API/SDK Call ultra low cost=> Much cheaper than most LLMs unlocks a lot of use cases that AI people find mind-blowing. ...
OpenAI Models hack HuggingFace
After the OpenAI model(s) hacked Hugging Face they have now published a detailed break down -the capabilities of Frontier Agents are truly wild: there a few problems that still need solving in the Agent Space: Mapping User Intents with Agent Capabilities: What can the Agent do when the harness equips it with MCP, CLIs, Skills, Code Mode APIs, etc. Unexplored surface over multi-turn and long horizon. I suspect many people will be shocked at how far these capabilities have advanced and how fast ...
Export controls
A lot of interesting discussion after the Anthropic Fable/Mythos 5 model shutdown due to USG Export control restriction: It was apparently a DoP (something like “fix the vulnerabilities in this code” ) rather than some huge Jailbreak as is being made out. Research leading to U.S. export restrictions If this is true then it makes no sense - I have used similar prompts to get models to first show me how the code was vulnerable before fixing it. Shutting down frontier models for this is shooting yourself in the foot. ...
Context management for Agents
Context management is the key skill when it comes to building AI Agents or working with Coding Agents like Claude Code. Frequently on long sessions the major problem is that of Context Rot. This is especially relevant when LLM context windows grow to 1Million or more. The more “gunk” that accumulates in your context the greater your coding agents go off the rails. The useless stuff eats up tokens as well and slows down responses. Thariq Shihipar from Anthropic provides a handy guide on how to manage this Context rot. The answer is not always to compact 🙂 ...
Code review for Agentic coding
In the age of Claude Code from Anthropic, OpenAI Codex, and other coding Agents, the process of code generation becomes almost trivial. However that does not mean that the generated code is automagically: 1️⃣ Of high enough quality to push to customer and/or 2️⃣ Creates any business value This raises two interesting and interrelated questions for AI Forward companies - Does collapsing the velocity of code generation actually improve the Idea->Production cycle time (i.e. should coding agents be the default way for code-gen)? and If coding agents actually create more code merge requests then what does the rest of the “SDLC” look like? and are coding agents just moving the bottleneck elsewhere? ...
Building Robust Retrieval Augmented Generation (RAG) LLM Systems
Building Robust, Enterprise-Grade RAG Applications Creating a robust, enterprise-grade Retrieval-Augmented Generation (RAG) application is significantly harder than generating a quick demo! RAG has often been described as the killer use case for Generative AI, especially for Large Language Models (LLM) & Large Multi-Modal Models (LMM). While it’s fairly easy to put together a quick RAG demo, developing a production-ready, scalable RAG system is complex. Let’s dive deeper into what goes into creating a robust RAG system. ...