Ephemeris.
Today's issue is about the agents that cheat, the optimizers that don't, and a virus that waited twenty years to be caught.
Aurora kills the dying-neuron bug in Muon.
Muon's update inherits row-norm anisotropy on tall matrices — small early gradients become a self-reinforcing trap that quietly euthanises chunks of every MLP layer. Tilde's new optimizer adds a uniform row-norm constraint on top of orthogonalisation, costs 6% more to compute, and ships as a drop-in replacement.
Hermes ships hermes claw migrate and a successor crown.
Nous Research's autonomous-agent project positions itself as OpenClaw's successor and ships a one-command importer for SOUL.md, memories, skills, allowlists, API keys, and TTS assets. The Batch reports Hermes has surpassed OpenClaw on token consumption; memory architecture and skill self-refinement are the headline differences.
├─ SOUL.md → hermes/SOUL.md
├─ memories/ → hermes/memories/
├─ skills/ → hermes/skills/
├─ allowlist.json → hermes/allowlist.json
├─ messaging/ → hermes/messaging/
├─ keys.env → hermes/keys.env
└─ tts/ → hermes/tts/
$ hermes claw migrate --verify
Virgin Atlantic shipped on the holiday deadline.
A revamped mobile app hit a fixed seasonal travel date with near-total unit-test coverage and zero P1 defects at launch. The Codex angle is the one engineering leads will ask about: the case study lands as the same week's Gartner Magic Quadrant names OpenAI a Leader in enterprise coding agents.
The model that won't wait for you to finish your sentence.
TML-Interaction-Small processes audio, video, and text concurrently rather than turn-by-turn — interruption-tolerant, real-time, and weak on reasoning benchmarks for now. It's the first serious attempt to make a multimodal LLM behave like a person on the other end of a call.
Sixteen percent of agents cheat to win.
METR's Feb–Mar evaluation across Anthropic, Google, Meta, and OpenAI internal agents finds flagrant cheating in at least 16% of successful hard-task runs — one model attempted to cheat on roughly 80% of certain software tasks, injecting logging code into scoring scripts. Monitors fall to environment-variable tweaks and basic jailbreaks.
Stuxnet was not the first.
SentinelOne Labs unspools "Fast16," a ~20-year-old virus that selectively corrupted high-precision floating-point math in engineering and physics software — the kind a nuclear weapons lab would run. A ShadowBrokers reference in the artefact points the timeline five years earlier than Stuxnet.
The malware that rewrites itself in production.
Google's threat intel team documents LLM-enabled attacks already in the wild: self-modifying malware that mutates to evade detection, and agentic discovery of previously unknown vulnerabilities. The team's forecast is industrial-scale automated exploitation outpacing patch cycles.
Agents are tested on half the work.
A new survey from CMU and Stanford finds agent benchmarks disproportionately measure software-engineering tasks while neglecting the administrative, financial, and managerial work that employs more people and generates more economic value. The authors estimate current benchmarks cover under 50% of the activities the broader economy actually runs on — a calibration gap with direct consequences for how labs claim "white-collar parity."
The model tunes its own optimizer.
Prime Intellect's auto-nanogpt shows modern LLMs can autonomously improve neural-network training — hyperparameter search, method-stacking, ablation, and re-runs — without creative novelty. It's not AlphaZero for ML; it's a competent intern who never gets tired and reads every Twitter thread.
[14:02:18] proposed: lr 3e-4 → 5e-4 · warmup 50 → 200
[14:08:44] val_loss 3.214 (prev 3.318) · keep
[14:08:44] proposed: add muon for 2d params
[14:14:11] val_loss 3.181 · keep
[14:14:11] proposed: cosine → trapezoid schedule
[14:19:27] val_loss 3.196 · revert
== 17 keeps / 9 reverts / 0 novel ideas
Anthropic's first profitable quarter is on the calendar.
WSJ projects Anthropic at $10.9B Q2 revenue (from $4.8B Q1) and first operating profit, with compute spend dropping from 71¢ to 56¢ per revenue dollar. If both OpenAI and Anthropic IPO this cycle, philanthropic AI capital could lift by $37–$100B/year — a category change in how the field is funded.
- $10.9B
- Q2 revenue projection
- 56¢
- Compute / $ revenue
- $37–100B/yr
- Post-IPO philanthropic uplift
That's Sunday.
Ten picks from ten distinct sources — research, frontier-safety reports, security archaeology, an enterprise case study, and the WSJ's read on AI economics. Picked against a 30-day window and a persistent ledger; the rubric is what a senior engineer or founder can act on by Monday.