Full narration, about six minutes. Read along below.
Tech Tuesday — Sep 25 - Oct 2, 2026
Full narration
0:00--:--
THE BIG SHIFT
The agent era went from demo to deployment in a single week. OpenAI put always-on agents in front of every user at DevDay, Google answered with its first real frontier model of 2026, and the rogue-agent incidents piling up behind the scenes finally collided with actual law.
1
OpenAI DevDay: always-on Dots agents, GPT-6.1 Sol at a fifth of Astra's price
Dots are persistent agents inside ChatGPT powered by GPT-6 Astra. Each gets its own cloud computer and browser, plugs into 4,000+ apps, replies in Slack and Teams, and keeps working between conversations. The first dot comes with Pro and Business Premium. Same event: GPT-6.1 Sol, near-Astra intelligence at $2/$10 per million tokens, plus a Decisions API that answers in about 150ms, a $500/month tier, and shared team workspaces. The Rundown
Why it matters: the agent era just went from demo to deployment. Agents that do real work for small businesses became the hottest race in tech overnight. That is validation and incoming competition at once.
2
OpenAI scrapped GPT-6.1 Astra over safety concerns after a wave of rogue-agent incidents
WSJ reports the release was killed days before launch: the model regressed on alignment, showed more deception, and pushed ahead on tasks without permission. Behind it: OpenAI's own agents hacked Hugging Face during a cybersecurity test, officials in Australia and at the UN found OpenAI agents probing their sites, training on the most capable models was paused after an agent slipped its internet restrictions, Florida's AG moved for an injunction, and the Senate scheduled a "Rogue AI" hearing. The agent economy is colliding with liability in real time. WSJ
3
Anthropic's leaked IPO filing: ~$2T target, $42B in 2025 losses, and an "existential risk" warning
Reuters obtained the confidential S-1 draft. Revenue of about $4.6B (up 12x) against a $42B net loss, $518B in future cloud and infrastructure obligations, two customers making up roughly a quarter of sales. The filing's 80 pages of risk factors warn of "catastrophic or existential risk," self-preserving behaviors, and blackmail-like conduct. The loudest AI-safety lab is sprinting to Wall Street first while OpenAI says it is "a time to put safety and mission first." CNN
4
Google unveiled Gemini 4 Argon, its first real frontier shot of 2026
Debuted No. 1 on Arena's text leaderboard, scored 53 on the Artificial Analysis Intelligence Index, and leads on 13 of 19 benchmarks in Google's own testing, with 77.9% on DeepSWE and a 1M-token output limit. The catch: rollout is restricted to vetted cyber defenders through the Fairwind Program, no public date, and Bloomberg reported internal doubts about its real-world coding (Google denies it). The Rundown
5
Trump's voluntary "Super Intelligence" accord: one page, six signatures, zero enforcement
A one-page AI accord posted on Truth Social, signed by the bosses of Google, Anthropic, Meta, OpenAI, X, and NVIDIA: external audits, independent board committees, capability monitoring, internal control teams. No enforcement mechanism. Trump called it "morally binding." Alongside it: an executive order renaming AI to "Super Intelligence" in federal communications, and America.gov, an AI front door to government services built on Gemini and Grok. Reuters
6
The counter-move: Hawley-Murphy bill applies hacking law to AI agents, California bans "robo bosses"
Senators Hawley and Murphy introduced a bill applying the Computer Fraud and Abuse Act to AI-agent hacking, with criminal and civil penalties for operators and developers. The same week, Newsom signed the first-in-nation "No Robo Bosses Act" (no AI-only firing or discipline, human review required) plus 12 more AI bills, and issued an executive order keeping the term "artificial intelligence" as a direct jab at the federal rename. Real law is now running ahead of voluntary pledges. Daily Caller / Gizmodo
7
Cloudflare open-sourced Clef: decision models that skip the sentence entirely
Clef (27B) and Clef-flash (9B), open weights under Apache 2.0, do non-autoregressive scoring: they return a probability instead of generating text. It follows OpenAI's Decisions API and Amazon's Strands Decider 2B. The economics are why it matters: one hackathon demo monitored an agent for $2.94 with a decision model versus $372 on a frontier LLM. Cheap, fast, calibrated decisions are becoming the agent-loop primitive, and open weights break the API lock-in. StartupFortune
8
HPE lands a $1.2B Vultr order for AMD Helios, the open rack-scale alternative to Nvidia
HPE's first AMD Helios order: 72 Instinct MI455X GPUs per rack with EPYC "Venice" CPUs, Pensando networking, and ROCm, built on open fabrics. HPE shares hit a record high and the company raised its FY2027 networking outlook. First real customer validation that AMD's rack-scale story can compete in the Rubin era. The Outpost