Channel
Digests from AI Engineer
LemonSlice discusses breaking the Avatar Turing test using world models. Key challenges include mitigating error accumulation over hours of video generation, optimizing for real-time performance, and building a robust model harness.
Read article
Todd Fisher demonstrates how he built a guitar plugin using JUCE and AI. He progresses from simple speech synthesis to complex conversational responses by integrating Whisper, local LLMs, and advanced DSP techniques.
Read article
Reelful builds an agentic video editor that automates complex content creation from raw footage. The system uses a multi-stage pipeline, running within a sandboxed environment and leveraging Remotion to express video as React code.
Read article
Krea 2’s training methodology reveals that data curation is everything. The talk covers advanced techniques like pHash deduplication, Sparse Autoencoders for unsupervised tagging, and multi-stage training from low to high resolution.
Read article
Code review must evolve beyond checking code diffs. The focus needs to shift from semantic accuracy (bugs) to preserving alignment, mentorship, and architectural intent by capturing session decisions.
Read article
Claw Patrol is an MIT licensed proxy that sits in front of your agents. It operates not at the HTTP level but at a lower level, understanding every byte flowing out of your agent to prevent destructive actions.
Read article
Context engineering is critical for LLM agents. The talk reveals that compaction techniques can be detrimental because they invalidate prompt caches, forcing full-price token usage. Full history retention often wins on recall.
Read article
The web is no longer just a source of data; we can look at it as a source of context for knowledge work using AI agents. This leads to Context-as-a-Service (CaaS).
Read article
RL worked when the world was a game, and IRL starts when the game fights back. The talk catalogues what a reward function meets on contact with a real login screen.
Read article
A CDP browser is just like a meat bag with a mouse. Your agent's clicks and keystrokes travel the exact same path inside Chrome that yours do. This requires a CLI, the CDP, and a Sense-Act-Verify loop.
Read article
Solving overhang is an engineering problem, and this is the work that we can do within our companies and within our agents... The most reliable browser agents that we see in production right now are often writing code alongside using the browser to actually automate a task.
Read article
Improving agents is a data mining problem. The key to continuous improvement is collecting comprehensive agent traces, which serve as the substrate for observability and continual learning.
Read article
This analysis explores how ChatGPT and Claude implement memory via running profiles. The speaker argues that memory is a function of compute and cannot be outsourced.
Read article
For modern AI agents, evaluation must shift from prompt playgrounds to robust code tests. Focus on 'raising the floor'—mitigating worst-case failures by tracking issue start dates and user impact.
Read article
Learn how Applied Compute brings continual learning to enterprises using a 2x2 distillation spectrum. Discover how online hinting and per-step guidance can boost agent performance from 15% to 80% without golden answers.
Read article
This guide details building LLM knowledge bases by transforming raw notes into structured wikis and graphs. The pipeline uses voice dictation for capture, agents for enrichment (tagging, linking), and cloud automation for scheduled updates.
Read article
AutoScientist automates model training by co-optimizing the entire loop from data to alignment, allowing builders to achieve frontier AI capabilities with greater efficiency and accessibility.
Read article
Yu Su argues that scaling AI requires moving beyond raw intelligence. True capability comes from 'Expertise,' which is achieved through continual learning to master specialized, complex domains.
Read article
Scaling Compute on Context addresses how to give LLMs deep expertise from private data when public scaling fails. It reviews methods like KV compaction and distillation to achieve self-improvement.
Read article
Memory harnesses solve context rot in long-horizon tasks by treating memory as a write-manage-read control loop, proving essential when answers fall outside the model's immediate context window.
Read article