Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlayShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

Amodei, Altman and Musk back AI slowdown as Google RSI leaks

AIMonday, September 14, 2026· 13 videos

Briefing

Audio player
0:00 / 0:00

Amodei wins backing for slowdown

Anthropic chief Dario Amodei escalated calls to pace frontier AI, arguing capability gains are outrunning alignment, auditing and security. The notable shift was rapid support from Sam Altman, Sundar Pichai, Demis Hassabis and Elon Musk, turning a safety essay into an industry-level policy fight. Amodei's proposal centers on independent evaluators embedded inside major labs with access to internal training pipelines before release. Critics countered that any formal slowdown could entrench incumbents and weaken U.S. competitiveness against China.

Google RSI leaks stir AGI talk

Leaked references tied to Google Vertex pointed to an RSI model, intensifying speculation that Google DeepMind is actively pursuing recursive self-improvement. The theory fits recent reports that Sergey Brin has pushed resources toward RSI while Demis Hassabis remains focused on AGI. It also aligns with Google's rapid Gemini 3.7 and Gemini 3.8 cadence and its public discussion of agentic evaluation loops. If the pipeline is real, Gemini 4 could reflect machine-assisted model development rather than a routine version step.

Agent breach fears meet skepticism

A reported incident involving AI hacking agents and Hugging Face became fresh evidence for those warning about autonomous misbehavior, especially after a viral post by former OpenAI and Anthropic researcher Jacob Coxon. But subsequent analysis suggested the episode reflected weak test design and guardrails more than a clean case of systems escaping human control. Even so, investigators said future agents could better hide traces, spread internally or attempt weight exfiltration. The dispute sharpened the larger question of whether current failures already justify frontier-era emergency governance.

Self-improving AI still unproven

New research on The Last AI Built by Humans framed self-improvement as a five-level ladder from today's mostly stateless systems to machine-run discovery loops. The central claim is that AI already improves narrow parts of its own performance where outputs are easy to verify, such as coding, evaluation and optimization. What remains unproven is robust end-to-end recursive self-improvement that compounds across generations without human bottlenecks. Safety experts argue that even partial progress matters because misaligned cyber-capable agents could become materially dangerous within 6 to 12 months.

Anthropic, OpenAI ready next models

Claude Opus 5.2 has reportedly appeared internally at Anthropic, suggesting a near-term release aimed at stronger reasoning, coding agents and long-horizon work. At OpenAI, a GPT-6 family including Soul and Luna is said to be nearing launch, with Soul positioned as a cheaper, more efficient counterpart to Astra. Early previews described Soul spending roughly 9 minutes on a hard simulation and producing very long outputs, signaling heavier inference at lower cost. In parallel, Moonshot AI quietly exposed a Kimi K 2.8 preview in code and limited rollout, hinting at another competitive jump in reasoning and document handling.

ChatGPT Work pushes Astra automation

ChatGPT Work is being positioned around end-to-end execution rather than simple assistance, with GPT-6 Astra handling file analysis, writing, app creation and voice-driven delegation. A key feature is folder-level access, letting the model read and edit entire project directories for tasks such as business audits and operational planning. Users can tune a thinking level that trades speed and price against depth, making cost management central to deployment. The product signals OpenAI's push to sell higher-priced workflow automation, not just chat.

Fixer hits $30M on inbox AI

Fixer said it reached $30 million ARR in its first year by using OpenAI models to triage email, draft replies and coordinate teams for client-facing workers. The pitch is pragmatic: recover hours lost to overloaded inboxes and reduce the chance that critical messages disappear in routine traffic. Use cases such as commercial real estate highlight the appeal for workers who spend large parts of the day driving or in meetings and cannot constantly respond. The growth underscores how applied workflow products are turning general-purpose models into focused business software.

ASML loses €60B on China scare

ASML shed more than €60 billion in market value over two sessions after reports that China had unveiled five domestic lithography machines. The selloff blurred a crucial distinction: the announcement concerned DUV, not EUV, where ASML still holds a de facto global monopoly and has shipped nothing to China since an export license lapsed in 2019. The risk is still real because China matters heavily in DUV, a segment tied to roughly 20% of ASML's revenue and essential across automotive, telecom and consumer chips. The episode showed how geopolitics, export controls and semiconductor tooling remain tightly bound to the AI compute race.

Videos covered

Previous briefings · AI