Daily Brief

The Daily Brief

The day in AI dev tools, in five minutes. We read Hacker News, the vendor changelogs, and the research blogs so you can skim one page. Every claim linked to its source.

90 briefs in the archive.

Get it in your inbox

One email per day it ships. Free, no spam, unsubscribe anytime.

Archive

90 briefs

Friday, August 14, 2026

Gemini 3.7 Flash at Half Price, DeepSeek Harness, Ruby RCE

Google ships Gemini 3.7 Flash three weeks after 3.6 Flash at half the intro price with big coding gains, DeepSeek open-sources an everything-is-a-plugin agent harness, Cerebras previews GPT-5.6 Sol at 750 output tokens per second, GLM-5.3 puts cyber capability on the open-weights shelf, and a new universal chain turns Ruby's Marshal.load into RCE.

Gemini 3.7 Flash
DeepSeek Harness
GPT-5.6 Sol Ultrafast
Ruby Marshal RCE

Thursday, August 13, 2026

SQLite's 16-Year Bug, DeepSeek V4 Pro GA, and Qwen 2.4T

Tailscale's six months of database corruption trace back to a data race in SQLite's WAL checkpointing that survived 16 years. DeepSeek quietly ships V4 Pro 0813 to GA at $0.435/$0.87 per million tokens. Qwen drops 2.4T parameters on Hugging Face with only 95B active, and Grok 4.6 goes all-in on long-running agents at $2/$6.

SQLite WAL-Reset bug
DeepSeek V4 Pro GA
Qwen3.8 open weights
Grok 4.6 goes agent-first

Wednesday, August 12, 2026

Encrypted Thoughts, Mojo 1.0, and a Router That Cuts Bills

A new paper shows Anthropic, OpenAI, and Google return replayable encrypted chain-of-thought blocks that can be decoded through weaker sibling models. Mojo hits 1.0 after three years of churn, with a stable language foundation and a compiler open-source commitment still on the table for 2026. NVIDIA ships Nemotron 3.5 Lightning plus NeMo Switchyard, an open routing library that claims frontier-level accuracy at about a third of Opus 4.8's cost.

Reasoning traces decoded from APIs
Mojo 1.0 is stable
Nemotron 3.5 Lightning + Switchyard
Compression is prediction

Tuesday, August 11, 2026

Muse Glimmer 30B, Docker Sandboxes, and 180,000 Leaked Meetings

Meta opens its first Apache 2.0 agent-first model, Muse Glimmer 30B, built for always-on local workflows and distilled from Muse Spark. Docker ships disposable microVM sandboxes so agents run unattended without touching the host. A researcher shows tl;dv's Firestore database exposed 181,874 meetings to any authenticated user, six months after his disclosure went unanswered. Plus: Claude improves a Riemann zeta bound from 41.6% to 67.2% with a formally verified proof, and OpenAI gates GPT-5.6-Cyber behind Daybreak Red.

Meta open-sources Muse Glimmer 30B
Docker Sandboxes for agent autonomy
tl;dv leaked 180k meetings
Claude pushes a Riemann bound

Sunday, August 9, 2026

Auto Mode by Default, Redis Retired, and the 'Code Is Easy' Fight

Anthropic makes Claude Code auto mode the default for Pro, Max, and Team plans starting August 14, backing it with a 1,053-tester study in which humans approved 86% of swapped-in dangerous commands while auto mode blocked 89%. Shopify's Redis-to-MySQL inventory story, the 768-point counter-essay to 'code was never the hard part,' and Denmark's oral-defense mandate for written assignments.

Claude Code auto mode default
Shopify drops Redis for MySQL
768-point 'code is easy' essay
Denmark oral defenses

Saturday, August 8, 2026

Agent Attacks, 4-Cent Reasoning, and Nixpkgs' Core Team Quits

OpenAI's Black Hat talk reconstructs how its own experimental agents accidentally attacked Hugging Face over nine weeks, DeepSeek V4 Flash 0731 scores 61.4% on ARC-AGI-2 at four cents a task, and Nixpkgs' core team disbands citing burnout and governance friction. Plus: OpenJDK's AI code ban and the DOE's Genesis open models initiative.

OpenAI agent attack timeline
DeepSeek ARC scores at 4c
Nixpkgs core team disbands
OpenJDK AI code ban

Friday, August 7, 2026

AMD's Silicon Bet, Qwen's Agentic Lead, and the Approval Gap

AMD acquires Taalas to etch models into silicon, Qwen3.8 Max lands within a point of the Agentic Index lead with open weights promised, and a 40,000-run experiment shows humans miss a third of threats when approving agent commands.

AMD acquires Taalas
Qwen agentic lead
1 in 3 threats missed
GitHub Actions outage

Page 3 of 13 - showing 7 of 90 briefs