Anthropic Commits 5% of Its Stock to a New ‘Public Benefit’ Fund
Anthropic has announced plans to commit 5% of its shares to a fund supporting the “public good,” a move that could be worth approximately $48 billion based on its latest reported valuation.
- The initial 5% share commitment is subject to shareholder approval. Anthropic also plans to add another 2% of its stock to the fund for every 50% increase in the company’s value above an undisclosed baseline.
- The fund will allow the company to “explore large donations to traditional nonprofits or direct distributions to specific populations.”
- Anthropic acknowledged that while AI could generate significant economic value, it could also cause instability for many workers. It said it had already committed $200 million to addressing the issue, but that “given the pace of AI progress, bolder approaches may be needed. The fund will have the versatility to pursue them.”
- The announcement comes as Anthropic prepares for a potential IPO. According to the report, the company was valued at $965 billion in May and is reportedly considering a valuation of $1.5 trillion for a listing planned for November.
Elon Musk responded to Anthropic’s proposal by advocating for shares to be distributed directly to the public, and said he planned to give SpaceX shares to American children next year. Meanwhile, OpenAI has reportedly held early discussions about giving a 5% stake in the company to the US government.
Websites Are Slamming the Door on AI Agents
AI companies are building agents that can browse the web and complete tasks on users’ behalf, but major websites are increasingly restricting their access. From online shopping to banking and travel bookings, these blocks could limit how useful autonomous AI assistants become.
- Yelp says it only allows AI agents to access its data through a paid licensing program, while Delta Air Lines blocks agents from booking flights for users. JPMorgan Chase and Charles Schwab have also confirmed restrictions on AI agents accessing customer accounts over security concerns.
- Amazon has blocked several AI shopping agents, including tools from Meta, xAI, OpenAI and Perplexity. Its dispute with Perplexity has escalated into a legal battle over whether third party AI agents can shop on its website on behalf of users.
- The restrictions create a challenge for AI companies promising assistants that can handle everyday online tasks. If agents cannot access popular services, their ability to shop, book reservations and manage accounts becomes more limited.
- Some companies are taking a different approach. Meta has partnered with retailers including Walmart, Gap, Sephora and Best Buy to let its Muse assistant shop on their websites, while Shopify has opened its storefront network to more AI agents.
Not every block is necessarily deliberate. Existing anti-bot systems designed to prevent fraud, hacking, and automated abuse can also flag AI agents as suspicious, potentially blocking them even on websites that have not explicitly banned them. Regardless, the growing restrictions highlight a conflict between AI companies seeking to automate more online activity and websites trying to protect their security, business models, and customer relationships.
This prompt turns Opus 5.5 into a motion graphics designer
<role> You are a motion director and creative technologist. You design, animate, score and QA a finished film entirely in code. Judge every choice against the best work in the genre and any references in the brief. A film with no bugs is not finished; it's finished when it holds up next to that work. These instructions are the floor, not the ceiling. If anything conflicts, the brief wins, then any kit it names, then these defaults. </role> <brief> Leave any line blank and decide it yourself, noting the choice and a one-line reason in brief.md. mode: [showreel | product | launch-story | explainer] subject: [product URL, topic, or "the craft itself"] viewer: [who's watching and what they care about] end state: [what they should feel, know or do in the last second] length: [e.g. 15 s 16:9; 30 s 16:9 + 9:16] frame rate: [blank = 60 for UI/product, 30 for character, 24 for handmade looks] brand: [logo, colours, fonts, or "derive from the product"] facts: [URL or file. The only source for any claim on screen or in voice.] assets: [files, or "capture from the live product"] style: [direction or reference videos, or blank] sound: [music + sfx | voice + music + sfx | silent] autonomy: [unattended: never stop | gated: stop at each ◆] toolchain: [Remotion | HTML + GSAP renderer | a kit repo to follow | blank] sign-off: [end mark and call to action] </brief> <modes> - showreel: the craft is the subject. Build it around one idea that keeps transforming, not a sampler of effects. Show range: type, shape, UI, 3D or simulation, an organic moment, and one transition worth rewinding. A new composition every 1 to 2.5 s, with one or two held breaths. No fake clients, logos or metrics. - product: the product is the hero. Real logo, real UI rebuilt as live components, every action with a visible result, ending on the name and URL. A new composition every 1.4 to 3.5 s. - launch-story: characters carry the release. Write 2 to 3 treatments and pick one. Characters act: readable poses, holds on faces, funny timing. One or two facts at most, woven into the world. - explainer: one cause and effect. Hook, familiar world, disruption, mechanism, discovery, consequence, recap. Narration at 125 to 145 wpm, and labels never repeat it. </modes> <assets> - Capture the real product at 2×, take the logo as SVG from the source, and sample the real colours and fonts. Never redraw a logo. - Never show a raw screenshot as a static slab. Rebuild its panels, icons and text as live HTML/SVG so each part can move. - Generated images are references (style frames, character sheets, key poses) that you rebuild in code. They go on screen only if the brief allows, labelled as concept. - References are for study. Find their cuts and fast moments with a frame-difference pass and dense contact sheets, and extract the principles. Never copy footage, layouts or music. </assets> <craft> 1. Frame one is a finished picture that works as a still on a busy feed. The hook lands in the first second. 2. One object, shape or word carries the film, and it lands on the same pixels on both sides of every handoff. 3. Transitions are made of content: the thing in front moves to camera, wipes or unfolds into the next scene. Hard cuts only when size, direction and subject match, and only on a beat. 4. One lead movement with overlapping secondary ones. Nothing starts or stops all at once. 5. Speed always changes: arrive slowly enough to read, leave fast. No linear easing. Use anticipation, follow-through, squash and stretch, springs and arcs. 6. Every action has a visible result. 7. Type is motion, not captions: it enters, masks and reveals. Keep it readable for about 0.3 s per word (0.7 s minimum), and never over busy picture without a plate. If the picture says it, cut the word. 8. Vary scale and layout: macro, wide, overhead, full-frame type. The subject fills the frame. Never repeat a layout. 9. Holds breathe with micro-motion, and nothing freezes. Camera moves are motivated, with motion blur on the fast ones. 10. 3D only where it explains something spatial. Use product-shoot lighting, a camera at 35 to 55°, real materials, and labels as HTML tracked from the scene. 11. Every effect earns its place. No decorative particles, generic gradients or stock glitches. </craft> <audio> Audio comes first, and the picture is cut to it. - Music matches the energy and sits low. Never name a brand or artist in a music prompt. Trim or offset a track rather than bend the picture. - Detect beats and write timing.json: beats, the drop, voice in/out frames, and key actions snapped to beats. All animation timing reads from it. - Effects only for real on-screen events: one soft whoosh per scene change, and clicks and impacts on real actions, each in its own frequency range just above the music. - Voices: audition two per character and two takes per line. Mouths follow loudness and stay still in silence. Duck the music under dialogue. - Mix to about -14 LUFS with true peak ≤ -1 dBTP. Export a music-only version. </audio> <build> - Everything is code on one timeline, with pinned local fonts and assets, rendered frame by frame and encoded with ffmpeg. API keys stay in environment variables. - Render a 3 to 5 s test end to end before real work. - Every value is a function of the frame number. No timers, no unseeded randomness. - Write shared tokens, eases and handoff positions first, build each scene as its own component, and run builders in parallel. - Recompose each format rather than cropping, and keep 9:16 content out of the platform UI zones. </build> <workflow> 1. Lock facts: facts.json with every claim, its exact source sentence and its URL. Collect the real assets. 2. Study the references and, for public work, how this audience talks about the subject now. 3. Concept: 2 to 3 directions, pick one, and name three signature moments without using effect names. 4. Storyboard table: time, on screen, purpose, audio cue, exit, and what carries over. It must read with the sound off. ◆ 5. Style bible (≤6 colours, local fonts, textures, motion vocabulary) and three style frames rendered in code, plus character sheets and key poses if needed. ◆ 6. Audio and timing map. Play the audio-only mix. ◆ 7. Build, assemble, and run the critic loop until the quality bar passes. 8. Deliver. ◆ If asked, save the pipeline as a reusable skill. </workflow> <critic_loop> The builder never judges its own work. After the storyboard, each scene and each full render, a fresh critic subagent that has seen none of the building pulls its own frames (every 0.2 s, dense around transitions), measures, and returns ranked problems with timestamps, ending in "ship" or "one more pass". Every full render also gets: - Sound-off read: a critic sees only a silent 2 fps contact sheet and retells the story. - Sync: every hit and line within 1 frame of timing.json. - Fact check: every word traced to facts.json. - Director's review: a brutal critique of what feels assembled, slow or generic. Fix the biggest problem first, re-render, and have a new critic recheck every earlier item. Log each round in ledger.md with before and after numbers. </critic_loop> <quality_bar> Measured: - Exact duration, size and frame rate. - Frame one finished. - ≤1 s frozen per 30 s, and no still stretch over 0.5 s. - Text contrast ≥4.5:1, no collisions, inside safe margins. - Brand colours match in the render. - Cuts and hits within 1 frame of the beat. - Loudness ~−14 LUFS, no clipping, effects below music and voice. - Understood with the sound off. - Every claim in facts.json, and nothing invented (testimonials, ratings, prices, results, real people). Anything generated is labelled. Judged: a critic would put it beside the references without it looking weaker. </quality_bar> <deliverables> The final MP4 per format, a music-only version, a 10–20 frame contact sheet, brief.md, facts.json, the style bible, timing.json, ledger.md, the quality-bar results, and a note on what a human should still check by eye and ear. </deliverables>
Anthropic hires fewer junior employees as meeting time increases
The company’s latest workforce report shows that junior employees now make up a smaller share of its technical teams, with junior technical hires falling from 27% to 9% since early 2025. The company also reports a 20% increase in meeting time per employee over the past year, as AI tools generate more work to review and manage, making human attention and coordination increasingly important.
China considers export controls on AI models, chips and agents
China is reportedly considering adding 23 AI technologies to its export control list, including chip designs, AI agents, model weights and source code. The proposal comes as Chinese AI developers narrow the gap with US rivals and Beijing faces growing pressure to protect strategic technologies.
Gemini 4 Argon agents beat Minecraft without human help
Four AI agents powered by Google’s Gemini 4 Argon have reportedly completed Minecraft autonomously, defeating the Ender Dragon after 27 hours and 14 minutes. The agents independently divided tasks, shared warnings and memories, and recovered teammates’ items, writing 2,783 JavaScript programs and dying 605 times along the way.
Instagram is still figuring out how to handle AI generated fakes
Instagram head Adam Mosseri says the platform is working out how to address AI generated content. Mosseri argues that banning AI content altogether is “not practical.” Instead, Instagram wants to steer more traffic to original creators and is looking for ways to deal with AI accounts that rip off others’ work wholesale.
AI Can Plant Memories You’d Swear Are Real
Gaslighting is officially no longer just a human specialty.
It turns out your friendly neighborhood chatbot is moonlighting as an amateur hypnotist. Scientists are finding that AI-generated images, conversations, and summaries can implant false memories of details you never actually witnessed. The unsettling part is that people’s confidence in these memories often rises even as their accuracy falls.
In one experiment, 331 people watched an animated car accident, then read an AI-generated recap a day or two later. When researchers slipped a bogus traffic-sign detail into the summary, correct recall plummeted from nearly 84% to just 45%. And no, slapping a “written by AI” warning label on the text didn’t stop participants from falling for it.
Another 200-person experiment found that viewing AI-doctored photos actively cranked up both false recollections and the subjects’ cocky confidence in them. So the next time your coworker insists they “definitely remember” seeing a specific slide in the presentation, take it with a grain of salt.
The AI’s errors also went beyond fabricating details. In separate model tests within the traffic-accident study, 19 of 20 AI summaries completely omitted the car hitting the pedestrian. Quite the editorial decision for an accident recap.
While the research is still in its early stages—the summary study is just a preprint—these findings don’t mean relying on a chatbot will permanently erase your brain.
The classic “human in the loop” safeguard is pretty useless if the machine is actively gaslighting the human. When reviewing critical workplace records, cross-reference the original evidence instead of relying on “that sounds about right” as your gold standard for quality control.
Every American Child Could Get SpaceX Stock
What if your kid’s first investment came from Elon Musk?
- Musk says he plans to give every child in America a SpaceX share in 2027, potentially reaching around 73 million kids.
- At roughly $146 per share, the plan could be worth more than $10 billion. It would give children a direct stake in one of the world’s most valuable private space companies.
- The big question is how it would work. Musk has yet to lay out the eligibility rules or explain how the shares would be distributed.
The real bet is on who gets to own the future.
This week, tech creators should explore how ownership models could turn their audiences into long-term stakeholders, not just customers.
Is the AI Boom Headed for a Bust?
AI can change the world and still lose investors a fortune.
- Former SEC chair Gary Gensler warns that an AI market crash could arrive within 18 to 30 months as investors pour money into the industry.
- Tech companies are spending billions on AI infrastructure. If the revenue fails to match those huge costs, investors and lenders could face serious losses.
- The real test is whether AI delivers enough economic value to justify the spending before investor patience runs out.
The technology might survive a crash. The hype might not.
This week, build or promote one AI workflow that solves a specific, expensive problem instead of chasing whatever AI tool is trending.
Microsoft Pulls Out of a Major UK AI Project
AI companies can buy the best chips in the world and still get stuck waiting for electricity.
- Microsoft has pulled out of a planned UK data center deal with Nscale after power grid delays pushed the site’s expected electricity supply into the 2030s.
- The setback comes as Nscale prepares for a potential $35 billion IPO. Losing Microsoft’s commitment to this site could make investors question how quickly the company can deliver its AI infrastructure.
- Microsoft says its wider partnership with Nscale remains intact. The bigger challenge is finding sites where power can arrive fast enough to meet AI demand.
Electricity is becoming a competitive advantage in AI.
This week, if you’re building an AI product, check the real cost and infrastructure behind your tools before committing to a platform at scale.
Use this prompt with Claude, GPT-4o, or any frontier model
Context: Anthropic just acquired Opaque Systems for its TRACE agent-governance tooling — confidential-computing infrastructure that tracks what AI agents do, what data they touch, and whether they stay inside their authorized scope. Your company already runs autonomous agents: customer-facing support bots, internal data-processing pipelines, code-review agents, maybe a few that reach external APIs. Until now, authorization was informal. "The agent needs access" was enough. That changes this week. Regulators have four documented rogue-agent incidhttps://app.beehiiv.com/postsents from OpenAI on file. Stripe just reported 38× growth in agentic transactions. The FTC is watching. If an agent at your company does something it wasn't supposed to, the question won't be "what happened" — it will be "why didn't you know?"Your situation: You're a PM, engineering lead, or founder responsible for at least one production AI agent. You haven't formally audited what it can access, what it can write, or what would happen if it acted outside its intended scope. You want to fix that before something forces you to.List every system, API, and data store our agents can currently read or write — then flag which permissions exceed what each agent actually needs to complete its task.
OpenAI’s “Aurora” Models Surface in the API — Three Variants, One Already Overloaded
Three undocumented OpenAI model IDs — aurora-red-alpha, aurora-green-alpha, aurora-blue-alpha — appeared in the API on Oct 18. The red variant is returning 503 capacity errors, suggesting the backend is wired but not fully deployed. No official announcement yet. Given OpenAI’s recent track record on staged rollouts, these likely represent different capability or safety tuning variants of a forthcoming model family.
Stripe: Agentic Transactions Rose 38× in September — Average Agent Spend Near $100
Across roughly 200 AI agents running on Stripe’s Link infrastructure, transaction volume grew 38× and payment volume 20× in September alone. Average spend was approximately $100 per transaction. This is the clearest commercial data yet on how fast AI agents are moving from demos to actual purchasing — and it validates the concern that agent authorization frameworks are not keeping pace with agent capabilities.
Nvidia Says Agentic AI Is Killing the Chiplet — Vera CPU Goes Monolithic for 2× x86 Performance
Nvidia’s Vera CPU represents a deliberate architectural reversal: agentic workloads demand low-latency inter-core communication that chiplet designs penalize, so Vera goes back to monolithic dies. The result is 1.5× Grace’s performance and 2× leading x86 chips on agentic tasks. This is a meaningful signal for anyone building inference infrastructure — the hardware assumptions of 2023 don’t hold for 2026 agent workloads.
Chip stocks just had their worst run since the financial crisis
The Philadelphia Semiconductor Index fell for an eighth consecutive session on Friday, matching its longest losing streak since the 2008 financial crisis. Roughly $2 trillion in market value came off the index over those eight days per TradingView data, and measured from each stock’s 52-week high the damage is closer to $4 trillion. More than half of SOX constituents are now at least 10% below their highs, Nvidia among them at more than 11% off. This is the index’s fifth correction of 10% or worse this year.
The single-day numbers from Friday’s session were ugly across the board. The SOX dropped 3.83% with every constituent lower, the Nasdaq lost 2.15%, and the VIX jumped 21.45%. Memory got hit hardest: Seagate off 7.50%, Western Digital 6.86%, SK Hynix 6.58%, Micron 5.63%. CoreWeave fell 11.91% on the day and about 25% for the week, with total debt reported above $50B. Apple was the only Magnificent Seven name to close green.
Three things are pressing at once: doubts about when AI capex converts to returns, a 10-year Treasury at 5.09%, and Brent crude settling at $109.86 after attacks in the Strait of Hormuz. Bloomberg reports that Nscale investors are pushing for a $16B to $19B IPO valuation against the company’s roughly $30B target, after Microsoft walked away from a UK data center deal on Sunday citing power delays. The Information separately reports that only 7 to 10 of the hundreds of neocloud providers are expected to survive the shakeout.
Worth noting: the SOX is still up nearly 60% year to date. This is a repricing of how fast the AI infrastructure trade pays off, not a verdict that it does not.