AI Daily Dev — November 5, 2026

NVIDIA Open-Sources Auto Ontology to Stop Agents Writing the Wrong SQL

Nov 1 release of a Python framework that builds a certified semantic graph over enterprise databases so agent-generated SQL carries approved definitions, relationships, and lineage.
Apache-2.0; connectors for Snowflake, Databricks, Postgres, DuckDB; ships a FastAPI backend, Next.js UI, pgvector catalog, and an MCP server exposing 9 read-only tools.
Vendor benchmark claims a ~90% reduction in ontology build time; accuracy scores are not published — the separate `ontology-sql-eval` repo ships only the harness.
Tops today's GitHub trending with +2,050 stars, ahead of claude-mem, worktrunk, and vercel-labs/agent-desktop.

open-source marktechpost.com

Oct 28 update broadens the Claude-credits program from individual researchers to funders (NSF, Sloan, Simons), institutions (HHMI Janelia, Allen Institute, Sanger, EMBL), and six research projects.
Partners redistribute credits internally or into existing grant pipelines; gnomAD and the Column Group are among the new additions.
Follows Anthropic's $150M three-year commitment to the federal Genesis Mission announced Oct 8.
Lands the same week as the Oct 27 drug-discovery progress report with BigHat, BMS, Divergence Labs, Enveda, Generate Biomedicines, and the Column Group.

research anthropic.com

Oct 27 early-progress report: Claude paired with BigHat–Anthropic teams to improve two antibody fragments — one targeting cancer cells, one extending a drug's half-life.
Covers active collaborations with Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group — all preclinical.
Anthropic quietly acquired Mithril Medicine, a one-person biotech working on two rare pediatric muscular dystrophies, for its in-house program.
Caveat baked into the post: real impact requires much deeper integration into partner development pipelines — this is a progress note, not a shipped drug.

research anthropic.com

SEC's 15-day rule after the Oct 26 S-1 means the earliest permitted roadshow start is Nov 10; press reports point to a Nov 9–mid-Nov kickoff and a target listing the week of Nov 16.
Public S-1 cover still reads 'between $ and $' — the price range lands in an S-1/A filed before the roadshow begins.
Target valuation: around $2T; secondary-market bids on Forge and Hiive sit near $1.05T with asks at $1.15T, versus the May private round at $965B.
Deal is expected to raise ~$60B, led by Goldman Sachs and Morgan Stanley; ticker ANT on the Nasdaq Global Select Market.

industry itiger.com

FareedKhan-dev/kimi-k3-in-c inferences the 2.78T-parameter Moonshot MoE on a single AVX2 CPU with a measured 8.24 GB peak RAM — no GPU, no BLAS, no framework.
Only 16 of 896 experts activate per token; the other 93% of the 1.56 TB checkpoint (1.447 TB) is read from NVMe per token and never enters memory.
Pure C99, Apache-2.0, ~176 KB of source; Linux x86-64 only, and reported throughput is about 32 seconds per token — a proof of concept, not a server.
Still crawling up GitHub trending three months after the Aug 1 release — a reminder that extreme MoE sparsity makes trillion-param inference a storage problem, not a compute one.

open-source github.com

max-sixty (PRQL, Xarray) ships a worktree manager pitched for running Claude Code, Codex, and Gemini CLI in parallel without stepping on each other.
`wt switch -c -x claude feat -- 'Add auth'` creates a branch + worktree and launches the agent in it; `wt list`/`wt remove` do the janitorial work.
Hooks on create/merge, LLM-generated commit messages, per-worktree dev-server port allocation, and CI status per branch.
+531 stars today on GitHub trending (now ~4.1K total); installs from Homebrew, Cargo, conda-forge, and winget.

tools github.com