Issue #84 2 min read

AI Engineering Signal #84

Kimi K3 open-weight MoE (2.8T params, 50B active) ships

Share

Signals

Kimi K3 open-weight MoE (2.8T params, 50B active) ships

Opus 4.x-class quality at Sonnet-tier pricing; Moonshot suspended new subscriptions under compute load.

Latent Space

Claude's Fable agent produced a Jacobian Conjecture counterexample

gate agentic math outputs through formal proof-checking before any citation or publication.

Web

HuggingFace security incident: attacker bypassed guardrails, forensics couldn't

rotate HF tokens and audit model artifact access permissions now.

Web

GPT-5 found a $500k WordPress RCE for $25 in API spend

automated exploit discovery is cheap enough to industrialize; revisit patch cadence and vuln triage SLAs.

Web

OpenAI cuts Codex context window from 372k to 272k tokens

pipelines sized to the prior limit will silently truncate; audit max-context assumptions in Codex workflows.

GitHub

Room-temperature quantum material demonstrated at LSU

cryogenic cooling requirement removed; watch qubit fabrication and procurement timelines for near-term shifts.

Web

Get signals like this in your inbox

Daily AI engineering intelligence. No noise.

[ Subscribe ]

The Take

The same week the largest open-weight model ever ships and a $25 API call finds a six-figure exploit, HuggingFace confirms attackers operate unconstrained by the guardrails slowing down defenders. Capability is outpacing the operational controls built around it — that gap is now the primary engineering risk.

Subscribe

Unsubscribe any time.

Related Signals