AI is starting to work on other AI. MIT researchers used OpenAI's GPT-6 Astra to edit a small model's weights directly, and its maths score jumped sixfold. The bills keep climbing too. Anthropic's amended IPO filing lists $895.4bn in compute and infrastructure commitments, Oracle is raising $20bn, and Marvell is paying $115bn for a laser maker. Anthropic is also paying up to $500,000 a year to bring senior people from other fields into AI safety work.
48.1%
a small Qwen model's maths score after GPT-6 Astra edited 0.016% of its weights, up from 7.9%
$895.4bn
in compute and infrastructure commitments in Anthropic's amended IPO filing
$115bn
Marvell's all-stock deal for AI data-centre laser maker Lumentum
IN THIS ISSUE · 5 SECTIONS
1 Major AI News: GPT-6 Astra edits another model's weights, Anthropic's paid safety residency, the AI build-out bill.
2 Fun AI Topics: Microsoft's live speech-rewriting patent, Claude's limits on game clones, a Ghibli-style short from one prompt.
3 AI Tools: DeepSeek Creator Mode, Cedric, Hera, UnleashX, Cyclon.
4 Expert Prompt: show your manager what AI delivered for you.
5 Trending: Claude's Message Threads, regret over AI layoffs, OpenAI's Abu Dhabi deals.
Forwarded this email? Subscribe here →
THE BIG STORIES
1 Major AI News
LEAD STORY
AI starts improving other AI without retraining
MIT researchers found that OpenAI's GPT-6 Astra can improve some models by directly rewriting the weights that govern their behaviour. Astra changed only 0.016% of a small Qwen model's weights and raised its maths score from 7.9% to 48.1%, while barely improving Google's Gemma 3, which lacked a similar isolated flaw. In a separate test, Astra wrote a small network from scratch that plays the game Breakout. The work has not been peer reviewed.
Why it matters: If models can reliably fix one another, some improvements may no longer need a full retraining run.
Astra worked from the weights alone, with no retraining run
Critics say the fix stopped the model looping rather than teaching it more maths
ALSO TODAY
Anthropic opens a paid AI safety residency for career switchers
Anthropic has launched a paid research residency for technically accomplished people who want to pivot into AI safety, with no prior machine-learning background required. Placements run from six weeks to six months and pay $350,000 to $500,000 annualised and prorated, with a dedicated mentor, GPU compute, access to Claude, and on-the-job training. Residents work on scoped projects in interpretability, AI control, scalable oversight, and red-teaming, with a path to publishing or joining full-time. Applications close on 22 November.
Why it matters: Anthropic is paying top-tier salaries to retrain experienced people from other fields, a sign of how short the supply of safety researchers is.
Professors, technical founders and senior engineers are the target, not just ML specialists
Visa sponsorship is available and short placements can be remote
ALSO TODAY
The AI build-out keeps getting more expensive
Anthropic's amended IPO filing lists $895.4bn in compute and infrastructure purchase and lease commitments, alongside nearly $17bn in third-quarter revenue and a $567m operating loss. Oracle is raising $20bn for its AI build-out, split evenly between stock and bonds, and its shares fell sharply on the news. Marvell separately agreed to buy AI data-centre laser maker Lumentum in a $115bn all-stock deal.
Why it matters: The scale of these commitments shows how much capital is being locked up to support AI, and how sensitive investors are to the spending.
Anthropic put its quarterly loss down mainly to about $5bn of short-term compute bought at premium prices
Oracle's share drop signals market nerves over AI capex
FOR THE COFFEE BREAK
2 Fun AI Topics
Microsoft patents AI that can rewrite live speech and faces
Microsoft has filed a patent for AI that can rewrite a speaker's words during a live video and reshape their lips and face so the edit looks natural. Agents would check speech against preset rules or viewer profiles, swap or delete flagged words, and patch the audio and video, with replacement language possibly pulled from a web search. Suggested uses include live translation, removing fillers, and filtering slang or offensive terms. It is a patent filing, not a product.
Why it is interesting: It raises the odd prospect of different viewers watching different versions of the same live event.
The original feed could be kept as a record
Microsoft declined to comment on the filing
Source: Patentlyze
Claude starts refusing game recreation tasks
Claude Code users report the model now refuses some game reverse-engineering and recreation tasks, including work using tools like Reverse Engineer Anything. Anthropic confirmed limits meant to stop Opus 5.5 from cloning specific games. Developers say the guardrail overreaches, blocking ordinary work such as copying a gameplay mechanic or patching an existing project, even after renaming it.
Why it is interesting: It shows how intellectual-property guardrails can spill into legitimate development and frustrate users.
Refusals appear even in routine coding work
Separate reverse-engineering restrictions remain unconfirmed
Source: Game Developer
A Studio Ghibli scene generated from a single prompt
A model reportedly generated an entire Studio Ghibli-style anime short from one detailed prompt, delivered as a self-contained HTML file with no external assets or separate image or video model. The scene shows a giant robot waking in a mossy forest and rising into a golden sky, produced procedurally through code. The file includes a custom player, fullscreen and sound controls, and a film-grain overlay, and runs offline in a browser.
Why it is interesting: The jump in quality between model versions shows how fast creative coding is improving.
No external dependencies are needed to run it
An earlier version produced a noticeably weaker result
TRY THIS WEEK
3 AI Tools
[PLUGINS] DeepSeek Creator Mode - Builds AI plugins from plain-language descriptions and adds them directly to DeepSeek Harness. deepseek.com (developer preview page)
[MEETINGS] Cedric - Joins Google Meet, Zoom, or Teams to present slides, take notes, or stand in for you from a brief. meet-cedric.com (product page)
[DESIGN] Hera - Generates editable graphics and motion designs on a single canvas. hera.video (product page)
[AGENTS] UnleashX - Offers ready-made workplace agents connected to services such as Salesforce and HubSpot. unleashx.ai (product page)
[PHONES] Cyclon - An AI operating system on a Google Pixel that uses existing apps and builds new ones on demand. cyclon.ai (pre-order page)
COPY AND PASTE
4 Expert Prompt
Context: A simple way to show a manager what AI actually delivered for you, not how much you used it. Gather your task list, meeting notes, or sent work first, leaving out anything your employer forbids putting into AI.
Review the material I paste below and identify each task where AI assisted me. For every task, list: what I asked AI to do, what I personally checked, edited, or decided, and the measurable result. Only include numbers that are real, not estimated. Then produce a one-page summary suitable for a performance review that highlights both outcomes and my own judgement.
Example use case: Run it every Friday to build an evidence log of AI-assisted work ready for review season.
WHAT PEOPLE ARE TALKING ABOUT
5 Trending AI Topics
Claude's API remembers conversations so agents answer faster
Anthropic's Message Threads, now in public beta, store conversations on Anthropic's servers so apps only send new messages plus the previous response ID, instead of the full history each time. On very long requests of 500,000 to 1 million tokens, the median wait for the first word of a reply fell by about 35%, and it is switched on by default for most Claude Code users. Threads are encrypted in memory, never written to disk, and expire five minutes after the latest request.
Why it is important: Long-running AI agents, such as coding assistants, start answering sooner without resending the whole conversation every turn.
Source: Anthropic
Layoff remorse grows as some firms rethink AI cuts
Palantir's CTO, Shyam Sankar, says almost all the customers he has seen pursue AI-driven headcount cuts failed to reach that goal, and a 2025 Orgvue survey found that 55% of firms that made staff redundant because of AI now regret it. In his view, the companies pulling ahead use AI to take on work that was previously too costly, letting small teams do far more.
Why it is important: Boardroom sentiment may be shifting from cutting jobs to expanding what existing teams can do.
Source: Stratechery
OpenAI expands in Abu Dhabi as UAE investors eye its funding round
OpenAI signed four agreements in Abu Dhabi with state oil firm ADNOC, the government, investor Mubadala, and G42's Inception42, as UAE investors are reported to be leading OpenAI's next funding round of up to $30bn. ADNOC will use decades of data from thousands of wells to cut oil-field planning from months to hours, while a 'machine-readable government' project turns food-import rules into text AI systems can check. A five-year deal with Inception42 covers AI adoption across government, business and education.
Why it is important: It makes Abu Dhabi both a major OpenAI customer and a potential top investor, deepening the ties between AI and state infrastructure.
Source: Abu Dhabi Media Office
That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas
