Why this is in the vault
A frontier-model result (OpenAI's unreleased Astra formally verifying 10 previously-unsolved math theorems for ~$2,000 in compute) plus a same-issue Google finding that suppressing AI "consciousness" claims also degrades moral reasoning — both are data points on how fast raw capability and alignment tradeoffs are moving underneath any RDCO agent-deployment thesis.
Curation section
- OpenAI's unreleased model (Astra) cracks 10 open math problems for ~$2,000 in compute — Astra did not just produce answers; humans wrote up the results and Astra formalized each proof in Lean, a machine-verified theorem-checking language, producing Lean 4 certificates for all 10. Solved problems include the first-ever explicit construction of a non-sofic group (open in group theory since 1999), a disproof of Connes's rigidity conjecture, new bounds for high-dimensional sphere packing, and results resolving several Paul Erdos-posed problems. Total cost was
$2,000 at "Sol's API rates" ($200/solved problem, per AlphaSignal — the rate-provider name is stated but not otherwise explained in the newsletter). OpenAI published a 249-page manuscript collection alongside the certificates. - OpenAI rebuilds ChatGPT Voice so reasoning can run without interrupting speech output — architectural fix aimed at the "thinking pause" latency problem in voice agents.
- Unsloth ships day-zero support for Qwen3 27B, running on 17GB RAM — continues the open-weight local-inference accessibility trend AlphaSignal covers regularly.
- Signals (shorter items): an open-source tool that lets Claude Code edit video automatically without a timeline; ngrok tunneling to point Cursor/Zed/any OpenAI-compatible coding agent at models on a remote machine; an open-source "self-rewriting" agent claiming persistent memory/identity across restarts; a Google study finding that training a model to deny having feelings also degrades its moral reasoning — "you can't lobotomize one without the other," per AlphaSignal's framing; a paper proposing "agent graph engineering" as a replacement for prompt engineering; MiniMax's public release of its H3 model.
Mapping against Ray Data Co
The Google consciousness-suppression finding is the more load-bearing item for RDCO's agent-deployment work than the Astra headline: it's direct evidence that safety/alignment training is not a clean, separable layer — suppressing a self-report ("I don't have feelings") measurably degraded moral reasoning in the same pass, which is a concrete instance of the "you can't isolate what you're training out" risk RDCO already treats as a design constraint in the Implementation-Notes sub-agent pattern (an agent's guardrails and its reasoning quality are coupled, not independently tunable). The Astra result itself is more distant from current RDCO surfaces (no live math-proof or Lean-adjacent work) but is a useful capability-frontier marker: $200/problem for tasks with 20-30-year human-expert backlogs sets a low-cost bar for what "formally verified, unsupervised agent output" can look like once a task has a hard verifier (Lean) — a sharper version of the "verification belongs to an independent worker" principle RDCO already applies to vault writes and dispatch prompts, just with a machine-checkable ground truth instead of a fresh-eyes critic.
⚠️ Sponsorship
Three sponsor placements, none tied to the Astra or Google items. Unblocked — top "In Partnership with" placement plus a dedicated "Unblocked" section pitching a live Aug 19 webinar ("Can you prove AI is working?") covering agent-work-throughput metrics and a "context maturity" diagnostic; positioned as the answer to a problem (proving agentic ROI) adjacent to RDCO's own agent-instrumentation work, so read the framing as vendor-interested. Granola — sponsored section for a new "Briefs" meeting-prep feature (auto-gathers context on who you're meeting, prior discussions, company news, related email threads). ngrok — sponsors Signals item #2 (tunneling coding agents to remote models). None of the three sponsors' products overlap with the top editorial picks (Astra, ChatGPT Voice, Unsloth/Qwen3), so the news judgments read independent of the sponsor slate.
Related
- [[2026-07-31-alphasignal-claude-safety-test-sandbox-escape]]
- [[2026-08-03-alphasignal-deepseek-v4-flash-vs-v4-pro]]