Jul 25, 2026

A developer packed a 28.9 million parameter language model onto an ESP32-S3 microcontroller — costing roughly $8 — that runs entirely on-device at about 9 tokens per second. The breakthrough uses Google’s Per-Layer Embeddings technique, originally from Gemma models, to keep only a tiny embedding table footprint while the rest of the parameters stay computed on-chip.

Separately, a startup chasing a free-tier AI product chose to self-host DeepSeek V4 Flash on AWS spot instances. The mixture-of-experts model holds 284B parameters but activates only 13B per token and ships quantized to MXFP4, shrinking disk load to 150–160 GB. The team runs it on Blackwell-generation RTX PRO 6000 GPUs via g7e.24xlarge instances, avoiding per-token API costs entirely.

On the agent coordination front, Loom introduces intent leases and git worktrees so multiple AI coding agents editing the same repository declare overlapping plans before making changes. Meanwhile, Raft’s engineering team discovered that AI agents lack peripheral vision and overhauled their notification system to precompute relevance instead of forcing agents to pay attention costs on every message. A solo-built Apache-2.0 prototype pushes guardrails to the kernel via eBPF-LSM and Z3, hooking syscalls like `execve` and `connect` in a cgroup-scoped manner so even prompt-injected agents cannot bypass them.

For AI forensics, the SAGA framework multiclass-attributes synthetic videos down to the exact generator model, version, and team, using a novel video transformer architecture. Epistemic Engine models codebases as justified beliefs, scanning AI-generated code for vulnerabilities and predicting which beliefs are about to collapse. Similarly, @ttsc/graph gives coding agents a TypeScript-compiler-built graph of declarations and relationships, slashing token costs by not scaling with repository size.

In a revisit of classic architecture talks, Martin Kleppmann’s Samza talk challenges the stateless-backend/shared-database pattern, proposing Apache Samza on Kafka’s distributed commit log to avoid whole-system shared-memory concurrency problems. Finally, a controlled experiment testing eleven Claude Code PPTX-generation skills found that only Anthropic’s official document skill and slides_maker produced genuinely native, editable OOXML tables and charts — the rest resorted to uneditable placeholder images.