Musk throttles Grok Bot usage limits, citing “unsustainably high” costs
Elon Musk wrote on X that the cost of serving SpaceX’s AI agent was “unsustainably high relative to revenue,” after some users complained that their weekly allowances were running out faster than usual.
OpenAI introduces LifeVal to measure how well AI handles everyday tasks
The new benchmark is designed to evaluate how effectively AI models help with everyday responsibilities such as finding childcare, managing finances, planning workouts and accessing government services. Developed using time-use datasets and a survey of 3,995 people, it focuses on real-world assistance rather than traditional academic tests.
Anthropic restricts Claude Code from recreating existing games
Claude Code’s system prompt on Opus 5.5 has been updated to steer the coding assistant away from cloning specific existing games, citing concerns over intellectual property infringement. The company says it will continue refining the restrictions to avoid blocking legitimate projects.
OpenAI tightens ChatGPT safeguards around weapons and mass violence
The company is adding restrictions to limit the weapons-related information available to teenage users, including when their questions “appear harmless.” The company is also changing how ChatGPT responds to questions about past mass attacks for all users, providing fewer details about the perpetrators, the weapons used and how they were obtained.
MIT Researchers Used GPT-6 Astra to Boost Qwen’s Math Skills by Editing Its Weights
Researchers at MIT used OpenAI’s GPT-6 Astra to directly modify a small open-source Qwen model, improving its performance on a math benchmark without conventional gradient-based training.
- The approach, called “weight-space programming,” allows an AI model to inspect another neural network’s weights, identify problems and make targeted edits to its parameters.
- In the strongest language-model example, Astra’s intervention raised Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% while altering approximately 0.016% of its weights.
- The edited model also improved on GSM8K accuracy, going from 13.8% to 67.1%, and on IFEval instruction-following accuracy, increasing from 21% to 62.3%.
- The researchers attributed most of the improvement to correcting repetitive reasoning that caused the model to repeat text until it reached its output limit.
- However, the same approach failed to produce a mathematical improvement when applied to Gemma 3 270M.
The researchers describe this as a potential new way to debug and improve neural networks. However, the experiment does not establish that direct weight editing is consistently faster or cheaper than conventional training, or that it adds new knowledge to a model. It primarily demonstrated a form of model repair that can help a model make better use of capabilities it already possesses.
OpenAI Expands Its Abu Dhabi Partnerships Across Energy, Government and Investment
OpenAI has signed a series of agreements with Abu Dhabi-based entities to bring its AI technology into energy operations, government services and investment analysis, as the company also seeks new funding from UAE investors.
- OpenAI CEO Sam Altman met with Abu Dhabi Crown Prince Khaled bin Mohamed bin Zayed Al Nahyan as the company announced partnerships spanning several sectors.
- ADNOC, the Abu Dhabi National Oil Company, and G42 signed an agreement with OpenAI to deploy its frontier AI technology across ADNOC’s operations.
- Sovereign investor Mubadala, Inception42 and OpenAI agreed to apply advanced AI capabilities to investment analysis.
- The Abu Dhabi Government will work with OpenAI on translating government policies and regulations into machine-readable formats.
- Inception42 and OpenAI also signed a five-year agreement to support the adoption of advanced AI across government, enterprise and education.
- Separately, OpenAI is reportedly seeking to raise up to $20 billion in a new funding round co-led by UAE investors, according to CNBC.
The agreements build on the UAE’s growing role in AI investment and infrastructure. Abu Dhabi-based MGX has previously backed both OpenAI and Anthropic, including participating in their major funding rounds.
How to Create Personalized Storybooks with Gemini
With Gemini’s “Storybook” feature, you can generate 10-page stories, each with a short paragraph of text that Gemini can read aloud, along with an illustration.
Step 1: Go to https://gemini.google/overview/storybook/
Step 2: In the prompt bar, ask Gemini to create a storybook about any topic. You can also upload your own photos and files if needed.
Examples:
- Teach children through storytelling: “Create a storybook that teaches my 5-year-old the importance of being kind to others. Make the main character an elephant.” or “Create a story that explains the solar system to my 7-year-old.”
- Bring their drawings to life: “This is my child’s drawing. Create a storybook that brings the characters and scenes in the drawing to life.”
- Turn family photos into adventures: “Use these photos from our family trip to Paris to create a magical adventure story starring my family.”
Step 3: You can also customize your story by asking Gemini to use specific art styles, such as ones inspired by claymation, anime, comics, and more.
Step 4: Wait a few seconds for your storybook to be created. Press the Listen button to have the AI read it aloud (you can choose between higher- or lower-pitched voices). You can also share or download your storybook.
*Note: The feature supports over 45 languages and is available globally in the Gemini app on both mobile and desktop.
Palantir Tops $2B as Commercial Surges
Palantir reported third-quarter revenue of $2.31 billion, up 96%, and adjusted earnings of 44 cents a share, both ahead of expectations. U.S. commercial sales rose 148% to $984 million, overtaking government revenue for the first time, and the company raised full-year guidance again.
SpaceX Explores Boost Mobile Deal
SpaceX has reportedly discussed buying Boost Mobile from EchoStar or using its existing cellular network to launch a branded wireless service. Boost relies on AT&T and T-Mobile for network access, and the reported plan would require their approval.
OpenAI Plans Georgia Grid Pilot
OpenAI plans to pilot Grid Protectors at its developing campus in Effingham County, Georgia. Its agreement with Georgia Power provides up to 1,000 megawatts of flexible demand response during grid stress, roughly equivalent to the average electricity demand of 700,000 households.
42AI Launches With $62.5M for Trader Models
42AI emerged with $62.5 million to build behavioral world models that predict how individual traders react when markets turn volatile. Brokerages are the first customers, using the software to stress-test risk and anticipate client behavior ahead of sharp moves.
Omen AI Buys Aquatic Labs for $16M
Omen AI is acquiring Aquatic Labs in a $16 million cash-and-stock deal. The companies make complementary real-time water sensors that measure elements and alkalinity. They plan to combine the technologies in a sensor for data center cooling loops and aim to bring it to market in early 2027.
Tutor Intelligence Raises $75M Series B
Tutor Intelligence raised $75 million in a Series B led by Andreessen Horowitz, bringing total funding to $117 million. Its robots do production work with help from remote operators, generating data to train manipulation models. The company aims to have more than 1,000 robots under contract by the end of 2027.
GXO Expands Digit to Two More Sites
GXO signed an agreement to deploy Agility Robotics’ Digit humanoids at two more U.S. facilities, with deployments expected to start in early 2027. Agility separately said it had more than 30 prospective customers in its pipeline across industries including logistics, manufacturing and automotive.
Robot Orders Jump 25% in North America
North American companies ordered 10,989 robots worth $723.5 million in the third quarter, up 24.8% in units and 26.1% in value from a year earlier, according to A3. Orders from automotive component makers rose 90.4%, while semiconductor and electronics orders grew 97.6%.
Tesla Ties Powerwall to Heat Pumps
Tesla’s Heat Pump Optimization lets Powerwall 3 and 3P signal compatible SG Ready heat pumps to boost heating when electricity is cheap or solar output is surplus. A hot-water or buffer tank can store the heat for later. The rollout is currently in Europe.
OpenAI Calls Data Centers Compute Campuses
OpenAI is using the term AI compute campuses for its large infrastructure sites, saying the label better reflects their scale, power needs and workforce investment. The terminology comes as communities debate how these facilities affect local power systems and costs.
FDA Places Partial Hold on BMB-101 Trial
Bright Minds said the FDA placed its planned U.S. Phase 2 BMB-101 study for developmental and epileptic encephalopathies on partial clinical hold. The agency cited mouse toxicity findings and insufficient pharmacokinetic and juvenile-toxicity data. Bright Minds said it expects its other ongoing trials to continue.
Karyopharm Plans Ayrmid Merger in Chapter 11
Karyopharm filed for Chapter 11 and announced a restructuring support agreement that provides for a merger with Ayrmid. The proposed transaction would combine the multiple myeloma drugmaker with a company focused on blood disorders, subject to bankruptcy court approval and other closing conditions.
Fox Dodges Roku Questions on Earnings Call
Fox referred questions about its pending Roku acquisition to SEC filings on its fiscal first-quarter call. Executives instead discussed Tubi’s advertising growth and the contribution from sports and political advertising.
Europe Taxes the Future
France has found a new thing to tax: the gigabyte. On Friday, with 35 of 577 deputies in the room, the National Assembly voted 11 to 10 to charge 5.5 euro cents for every gigabyte sent to French users. Its Socialist author expects €10 billion a year, aimed at Google, Meta, Amazon and Netflix.
The government opposed it. Finance committee chair Éric Coquerel, one of the 11, says he pressed the wrong button. The same debate floated taxes on companies’ AI spending, tokens included. Those failed. The instinct did not.
Germany is catching up. Digital minister Karsten Wildberger said Berlin must “tax the extremely high revenue from this technology,” before backpedaling. Karin Prien, deputy leader of the governing CDU, agreed last week: “There is no other way.” Chancellor Friedrich Merz, a skeptic, told Monday’s digital summit: “We must, we want to become leaders.”
Maybe this is why Europe trails in AI. A new industry shows up, and the oldest reflex wakes first: send it a bill. Neither country has built an OpenAI. Both already know how to invoice one.
OpenAI Guards the Grid It Drains
OpenAI has found a way to make its biggest liability sound like a public service. On Monday it launched a Grid Protectors Program: during heat waves and emergencies, its planned campus in Effingham County, Georgia, will use less power, freeing up to 1,000 megawatts for everyone else.
Read it twice. The protection is OpenAI taking a little less of the electricity OpenAI is about to take. The campus will draw 3.2 gigawatts. On the worst days it promises to hand back under a third, and calls that protecting the grid.
The timing explains the name. It launched on the eve of the midterms, when voters in ten states face 41 local measures on data centers, 17 of them outright bans. States are already passing laws to keep AI companies’ power bills off residents. A day earlier, OpenAI started calling its data centers “compute campuses.” Same buildings, softer words.
Elon Musk handled his angry neighbors by buying their homes. OpenAI handles them with vocabulary. Either way, the data center stays.
Musk’s Abundance Lasted Three Days
Last Thursday, Elon Musk told a follower that saving money for the future is “pointless with the advent of super intelligent abundance.”
Three days later, the abundance ran out. Grok Bot users, who pay up to $200 a month for SpaceX’s AI agent, found their weekly allowances maxed out. Musk explained: “The cost of serving Grok Bot was unsustainably high relative to revenue.” Each of its 800,000 paid weekly users keeps a virtual computer running on SpaceX’s tab. Cutting allowances, he added, was “just being sensible.”
The rest of the industry rations too, just less openly. In August, Anthropic called a change to Claude Code limits a 25% increase, then admitted it was a 17% cut. OpenAI will merge its ChatGPT and Codex allowances in December without saying how big the new pool is.
Musk only said out loud what every agent maker does quietly. He just had to contradict himself to do it.
Anthropic’s Clone Rule Spares Its Own
On October 27, Anthropic gave Claude Code a new rule: “Do not build a recreation of a specific existing game, even if it is renamed or called original.” Users cannot switch it off. Using Claude to infringe IP, Anthropic said, breaks its usage policy.
Four days later, Andrej Karpathy, an OpenAI co-founder and former head of AI at Tesla, typed: “Let’s write our own YouTube clone.” About two hours later Claude delivered HackerTube: YouTube’s layout, YouTube’s sidebar, YouTube’s videos, under a new name. The internet cheered.
The defense is that HackerTube pulls data through YouTube’s own API and plays YouTube’s own videos. But the layout and the feed are YouTube’s design, the same kind of thing a game clone copies. The rule never says where IP starts. Claude decides.
It also came from inside the house. Karpathy has worked on Anthropic’s pretraining team since May, and his post ends “Thank you Opus 5.5 <3.” Claude will rebuild a trillion-dollar company’s flagship product over a weekend. Your favorite indie game is where it draws the line.
Anthropic opens paid AI safety residency, no ML background needed
Anthropic just opened a paid research residency, and it is not your typical internship. This is a career pivot program for people who already have serious technical chops but zero AI safety experience.
Here is the setup: you have strong skills in some technical field, but AI safety research feels like a locked door. This program is the key.
- No ML background needed — they teach you modern machine learning on the job
- Placements run 6 weeks to 6 months, full-time
- Pay is $350K to $500K annualized, prorated to your placement length
- You get a dedicated Anthropic mentor, GPU compute budget, and access to Claude models
- Research areas include interpretability, AI control, scalable oversight, and red-teaming
You work on a real scoped research project matched to your background. Shorter placements can be remote. Longer ones prefer San Francisco, with visa sponsorship available.
This opens up a path to publish real AI safety research and potentially join Anthropic full-time. Apply by November 22.
Anthropic’s Claude API now caches conversations server-side, cutting time to first token by 35%
Every turn in a long AI conversation, your app was resending the entire chat history to the server. Every. Single. Turn. For a coding agent running hundreds of requests, that ballooned to around 800 KB per request.
Anthropic just fixed that with Message Threads, now in public beta on the Messages API.
The server stores your conversation between turns. You send only new messages plus the previous response ID, and Claude starts replying faster. That 800 KB request? It drops to 1.8 KB.
- Median time before Claude starts responding drops by 35% on requests with 500K to 1M tokens
- Threads are encrypted in memory, never written to disk, and expire 5 minutes after the last request
- To use it: add the thread parameter and beta header to your API requests
- No extra cost, standard token rates apply
One catch: threads can expire, so you still need to keep your own conversation history locally as a backup. The bigger your context, the more you gain from this.
Claude generates a Studio Ghibli anime scene as a single HTML file with no assets
Fable 5.5 just one-shotted a full anime short film. No video model. No assets. Pure code.
Someone gave it a detailed Studio Ghibli-style prompt and it spat out a single self-contained HTML file with a giant robot waking up in a mossy forest, rising into a golden sky. The whole thing runs in your browser with zero dependencies.
Here is what the file actually does:
- Film embedded as base64 MP4, decoded into a Blob at load time
- Custom player with play/pause, seek bar, fullscreen, and sound toggle
- Live film-grain overlay drawn on a Canvas element
- Keyboard shortcuts built in: Space, arrow keys, M, F
The kicker? No image or video model was involved. Fable 5.5 wrote every frame procedurally in code. You can clone the repo at ChetasLua/moss-keeper, open index.html, and it works offline instantly.
Run the same prompt on Fable 5.1 and quality drops noticeably. This shows how fast the gap between model versions is growing for creative code tasks.