"ChatGPT Work Agent login access, Claude Opus 5 FPS game, Nvidia motion accuracy, Swiss 262K context model" — AlphaSignal
Why this is in the vault
ChatGPT's Work agent now handles password-walled sites by pausing for manual credential entry (agent "sees nothing," no credential storage) and persisting the login across sessions — a concrete instance of the human-gated-action boundary RDCO already runs operationally (send-button human-gated for email, paper-trade deploy needs explicit authorization). Claude Opus 5 one-prompting a 55,000-line, 11-subsystem FPS game is a fresh capability-frontier data point for the L5 north star's "bets are downstream of agent capability" thesis.
Mapping against Ray Data Co
The ChatGPT Work login-persistence pattern is structurally identical to RDCO's own gated-autonomy design: the agent stops at a trust boundary, hands control to the human for the sensitive action, then resumes — same shape as Ray's "drafts, founder sends" email rule and the paper-trade deploy-verb gate. OpenAI shipping this as a first-class product feature (not a one-off skill) validates that pattern as an industry-converging default, not an RDCO idiosyncrasy — useful as an outside reference point next time the founder questions why an autonomous step needs a human handoff.
Claude Opus 5's FPS build is more a capability-tracking data point than an immediate operational input: AlphaSignal's own coverage discloses a critic score of 5/10 against real COD footage ("looks procedural up close"), so treat the headline as directional evidence of code-generation ceiling rising, not a claim that Opus 5 now produces shippable game assets.
Curation section
Top features:
- ChatGPT Work agent — login persistence (4,339 likes): agent pauses at login walls (Jira, Notion, internal dashboards named), user types credentials manually, agent doesn't see/store them, session persists after — one login per site. Users can clear saved logins in ChatGPT data settings.
- Claude Opus 5 — 55,000-line FPS game (6,652 likes): single prompt, in-browser Three.js, Call-of-Duty-quality target, 11 subsystems (rendering, physics, AI soldiers, weapons, audio, UI), no external art/sound assets — all code-generated at load, ragdoll physics + bullet penetration, 120x120m map. Self-disclosed critic score 5/10 vs real COD.
- Airi (3,223 likes): open-source alternative to "Neuro-sama" — animated AI companion, <500ms voice latency, plays Minecraft/Factorio autonomously, Discord/Telegram integration, 30+ LLM backends, MIT licensed, 17.5k GitHub stars.
Signals (one-liners, no expanded body — flagged as thin):
- Tool converting any book into a token-efficient Claude skill (3,210 likes)
- OpenDCAI's DataFlow — open-source LLM training-data pipeline toolkit (6,993 stars)
- Nvidia GPT-style model, claimed 99.98% accuracy generating human movement (1,318 likes) — no model name, paper, or team beyond "Nvidia" given
- Swiss AI open multilingual model, 262K token context window (1,427 downloads) — no model name or parameter count given
- Claude Opus 5 on ARC-AGI-3: 30.2% score, described as a new record (2,997 likes)
Note on sourcing: all outbound links in this issue route through AlphaSignal's own tracking redirector (app.alphasignal.ai/c?...), so true third-party destination domains aren't recoverable from the raw body without following live links. Zero deep-fetches performed this issue — per curation-mode rules, deep-fetches require a resolvable third-party domain plus a specific hook; the redirector layer and the thinness of the two lowest-detail Signals items (Nvidia, Swiss model — no model/paper names given) meant neither cleared the bar for a live follow.
⚠️ Sponsorship
Two paid placements, both clean third-party blocks with no disclosed AlphaSignal ownership stake. Bright Data is the issue-level "In Partnership with" sponsor, pitching a "Video Layer for Physical AI Training" — pre-cut MP4 clips with timestamps/metadata for VLA (vision-language-action) model training, CEO Or Lenchner quoted re: an event called "Machina." HydraDB sponsors a separate signal pitching agent context graphs — connecting Slack/Notion/GitHub/100+ tools into "company brains, agent memory, ontologies." Neither sponsor's product is presented as editorial coverage; both are clearly marked ad blocks. AlphaSignal's own self-promo (ad-partnership pitch signed "John, Director of Brand Partnerships," "320,000+ Developers and CTOs," footer "WORK WITH US" CTA) is house content, not third-party sponsorship.
Related
- [[2026-07-10-alphasignal-chatgpt-work-agent-claude-reflect-swe17]] — the original ChatGPT Work agent launch coverage; this issue is a feature-update follow-on (login persistence) to that agent
- [[2026-07-24-alphasignal-chatgpt-voice-claude-opus-managed-agents]] — same sender, same week's run of Claude Opus capability coverage (voice mode, managed agents) that this issue's FPS-game and ARC-AGI-3 items extend