This week's briefing (9–15 Nov)Data updated 6:35 IST, 22 NovHow we check claims
Andrew Ng
DeepLearning.AI / AI Fund · Researchers & skeptics. 2 sourced entries, newest first. Curated coverage with primary sources, not a complete feed; a role establishes context, not authority.
DeepLearning.AI / AI Fund · 16 October 2026 · Blog post
The Batch letter: don't overgeneralize from AI's progress on math and code
Primary source, read in fullModelsAI agents
Ng distinguishes verifiable tasks (coding, math) from judgment tasks (customer service, strategy) that lack ground-truth answers, arguing progress on verifiable tasks has been much faster because reinforcement learning and automated labeling work better there. Citing Deep Blue, Watson and AlphaGo as precedents, he warns that 'tremendous progress on verifiable tasks... is leading people to overgeneralize', and concludes substantial work remains on judgment-based applications where humans keep contextual advantages.
What this could change
Editorial: a practical rule for scoping AI projects — expect benchmark-level performance on verifiable work, not on judgment calls.
Perspective & interests
Ng sells AI education and invests in AI startups via AI Fund, so calibrated expectations that keep companies building (rather than disillusioned) support his businesses.
Reviewed 2026-11-06. Entry ID andrew-ng-verifiable-vs-judgement-2026-10 · Permalink
DeepLearning.AI / AI Fund · 2 October 2026 · Blog post
The Batch letter: AI's cybersecurity abilities should encourage us
Primary source, read in fullAI safetySecurity
Ng argues GLM-5.3's high ExploitBench score (12% successful exploit attempts versus Claude Mythos' 14%) is good news, because capable cyberdefense tooling is becoming broadly accessible. He notes $20.40 of GLM-5.3-Flash tokens was enough to find a recently disclosed Chrome flaw, and frames falling vulnerability-discovery costs as an engineering challenge like aviation safety, contending defenders ultimately hold the advantage because they have more information about their own systems.
What this could change
Editorial: the direct optimist counterpoint to the Hinton-Bengio containment letter, with concrete cost numbers on AI-found vulnerabilities.
Perspective & interests
Ng's education and venture businesses benefit from an optimistic, build-focused framing of AI risk.
Reviewed 2026-11-06. Entry ID andrew-ng-ai-cybersecurity-encouraged-2026-10 · Permalink