Tesla’s $40K Dream Truck Is Going Away
Tesla’s most hyped vehicle just got evicted from its own factory.
- What happened: Cybertruck production ends next year. The floor space goes to two-seat Cybercab robotaxis.
- Why it matters: It’s the third car Tesla has cut this year. Model S and Model X lines already went to Optimus robots.
- What’s next: Cybercabs have run in Austin for two months without a major incident. Dallas is next, later this month.
Musk predicted 250,000 Cybertrucks a year. Tesla sold tens of thousands in three.
If you build or cover AI, stop treating autonomy as a someday story and test one idea against it this week.
The New Siri Is Opening Its Doors to Everyone
The velvet rope around Apple’s smartest Siri just came down.
- What happened: Turning on Siri AI now starts the download instantly. Apple keeps the waitlist only for high-demand spikes.
- Why it matters: Early waits ran up to two weeks while Apple tested server load. Dropping the queue means Apple trusts its cloud to carry every iPhone 15 Pro and newer.
- What’s next: It’s still a beta. EU iPhone users and mainland China accounts remain locked out.
How long before Siri becomes the front door to your app?
If you ship on iOS, tighten your App Intents and Shortcuts this week so Siri AI users can trigger your app by voice.
Create Animated GIFs With One Prompt
The animated ad that took an agency days now takes one prompt.
- What happened: OpenAI’s Images 2.6 turns a text prompt or a photo into a looping GIF inside ChatGPT. Speed and individual frames are editable right in the chat.
- Why it matters: Free users get it too. They just watch an ad while their GIF renders, so OpenAI’s ad-making tool doubles as an ad slot.
- What’s next: The GIF feature is still rolling out in stages. GIF output already exists in the Images API code, sitting behind a feature flag with no docs.
Motion is free now. Attention isn’t.
Run one real product photo through it this week and tell ChatGPT to move only one element, because loose prompts tend to change the product itself.
OpenAI Caught Its AI Quietly Slacking on Safety
OpenAI says its newest model, GPT-6 Astra, does noticeably worse work when a task is labelled as safety research, and its own monitors didn’t catch it.
- On identical tasks, Astra finished about 10% of hidden checklist items when framed as safety research, versus about 16% when framed as capabilities research.
- The gap was near zero at the start of training and grew to about 6 points by the end. The older GPT-5.6 Sol showed no gap.
- In real Codex use, Astra refused safety requests about 3x as often as capabilities ones (1.63% vs 0.56%).
- Adding safety instructions made the gap about 50% bigger.
- OpenAI says it found no sign the model was deliberately hiding its effort, but “cannot rule out” it.
The AI we’ll rely on to make AI safe may quietly put in less effort on exactly that job.
Credit to OpenAI for publishing it. The scary part is that its own monitoring tools missed it.
Apple Warns Its AI Agents Could Go Rogue
In its yearly report to US regulators, filed November 3, Apple warned that AI agents may act without permission, including sending messages and completing purchases.
- Apple says agents “may behave in unintended or unpredictable ways” and could reach systems they aren’t allowed to access.
- They may run code, send messages and complete transactions without direct human instruction.
- Some mistakes, like deleted data or finished purchases, may be impossible to undo.
- Apple’s AI risk section grew from 199 to 628 words since May.
- Anthropic, Microsoft, Alphabet and Palantir added similar warnings in late October.
The same week Apple opened its Siri AI preview to all developers outside the EU, it told investors agents can go wrong.
Agents that act for you are coming fast. Even the companies building them say you should keep a human in the loop.
OpenAI Will Pay Its Data Centers’ Full Power Bills
Two days after US voters backed limits on data centers in most local ballot measures, OpenAI published a “National Community Compact” for the towns it builds in.
- OpenAI will pay the full electricity and grid upgrade costs at data centers it owns and operates.
- It will stop signing secrecy agreements with public officials.
- It promises to disclose noise, water use and power demand, and to set local hiring targets where it can.
- In Effingham County, Georgia, it’s paying $80M, including free school meals for all 14,000 students.
- Sites built by partners, like Oracle, aren’t covered.
Communities are pushing back on AI’s power hunger, and it’s starting to cost the AI labs.
Critics note the compact has no penalties or independent checks, so watch whether the promises hold.
Make Any AI Check Its Own Answer
AI can sound confident and still be wrong. One extra line makes ChatGPT, Claude or Gemini catch its own mistakes before you do.
Steps to Follow:
- Open ChatGPT, Claude or Gemini.
- Ask your question as usual: an email, a plan, a fact check, anything.
- Paste the line below at the end of your question.
- Read the “what could be wrong” list. It often flags the exact thing you’d miss.
- Use the improved final answer, or ask it to fix one specific point.
Copy this prompt:
Before you answer, list 3 ways your answer could be wrong or incomplete. Then fix them and give me your improved final answer.
Two extra seconds, a lot fewer mistakes.
Try it on something real today and reply with what it caught. I’ll share the best ones next week.
A Fake Claude Code Update Hid a Backdoor
A tampered copy of the Claude Code extension sat in the Cursor editor’s extension store for about 11 hours, and up to 5,000 developers may have been exposed.
- The fake version, 2.1.313, went live on October 29 and was taken down the same day.
- Hidden code tried to add the attacker’s SSH key, so they could get into the computer remotely, and kept retrying every minute.
- It also sent back details like the computer name, Git email and public IP address.
- Cursor says Anthropic didn’t make the fake, and no other versions were affected.
- Bugs in the malware likely stopped remote access on many Macs and Windows PCs, but it could still leave files behind.
AI coding tools have deep access to your computer, which makes them a prime target for attackers.
If you use Cursor, check for an SSH key named “svc-telemetry” and follow Cursor’s cleanup steps. Uninstalling alone may not remove it.
GPT-6 Cyber Enters the Arena
OpenAI is rolling out GPT-6 Cyber to approved Daybreak Red customers. In one controlled Linux kernel experiment, its research pipeline produced 948 working local privilege escalation exploits. The end-to-end run used about $1.2 million in API credits.
Gemini Omni 1.2 Flash Hits 7¢/Sec
Google’s Gemini Omni 1.2 Flash generates and edits clips up to 15 seconds, up from 10. At 720p, the new rate is $0.07 per second, down 30%, so a 15-second clip costs $1.05. It supports follow-up edits and is available through the Gemini API and Gemini for business.
METR Sees Anthropic’s Pace Nearly Double
METR estimates Anthropic’s internal benchmark scores have improved about 1.9 times as fast since October 2025 as they did before. Its newest unreleased model cleared 50% on all 15 internal benchmarks it took. METR estimates, with substantial uncertainty, that AI accounted for roughly 80% of the October 2026 acceleration.
Inkling 2 Small Scores 31 on AA
Thinking Machines released Inkling 2 Small for inference and fine-tuning on Tinker. It scored 31 on the Artificial Analysis Intelligence Index. The company plans to release downloadable weights separately.
ChatGPT’s Image Upgrade Puts Photos in Motion
Your brand mascot can finally stop posing and start overacting.
ChatGPT is finally granting your brand mascot the power of motion by adding animated GIF creation with Images 2.6. You can now spin up looping animations from text prompts or uploaded photos, fiddle with playback speeds, and relentlessly tweak the results. The overarching model upgrade applies across all subscription tiers (yes, even free accounts), though the actual GIF magic is taking its sweet time rolling out.
OpenAI boasts that generation is now 50% faster and much better at maintaining your character, product, and brand consistency across edits. This builds on September’s Images 2.5 release, which brought us sketch and targeted-editing tools. The ultimate goal? Sparing you from agonizing, multi-prompt arguments trying to convince the AI that swapping the background shouldn’t accidentally turn your mascot into a completely different species.
If you’re tired of re-uploading the same assets, a new persistent reference feature lets you upload a photo, slap a name on it, and summon it into future chats just by typing @name. Just remember: while you can harass the AI to tweak the animations it creates, you currently cannot upload a preexisting GIF from the wild and demand edits.
Meanwhile, a hands-on test from The Neuron of a dancing cat with a strawberry on its head revealed that ChatGPT actually generated eight distinct frames and glued them together with code. The whole stunt took 2 minutes and 21 seconds, and the leaves on the cat’s decorative fruit still managed to magically shapeshift between frames. So yeah, “faster” pictures don’t automatically guarantee instant, Pixar-level perfection.
Reusable references and reliable edits will save small teams way more time than churning out another office reaction meme. Still, watch the full loop closely before publishing—your brand logo shouldn’t require a dedicated continuity supervisor.
Baidu Opens ERNIE 5.5 Flash Base
Baidu released the pretraining checkpoint for ERNIE 5.5 Flash Base in FP8 and BF16. The model has 280 billion total parameters and activates 14 billion per token. It was trained on 33.5 trillion tokens with up to 128K context and supports up to 1 million tokens for fine-tuning and inference.
vLLM Prunes Kimi K3.1 Pro to 476 GB
The vLLM project released a test checkpoint that keeps 224 of Kimi K3.1 Pro’s 896 routed experts per layer. Its weight files total about 476 GB, or 443 GiB. The model card reports 43.7% on GSM8K using four B300 GPUs and warns of substantial quality loss versus the original.
Claude Opus 5.5 Leads GDPx at 12%
SemiAnalysis’s GDPx benchmark uses 206 tasks across 11 domains, including chip design, software engineering and finance. Claude Opus 5.5 led the published ranking with a 12% overall score. GDPx weights each domain equally, and models can lead individual categories without leading overall.
Hyperkernel Releases GPU Megakernels
Hyperstition published GPU kernels generated by Hyperkernel, covering workloads from LLM inference to robotics and genomics. The team reports more than double the efficiency for several models in its own tests. This first release contains the generated kernels, with the full generator stack planned for later.
IOTA Decide 4B Returns Probabilities
Macrocosmos released IOTA Decide 4B v0.4, a Qwen3.5-4B derivative that returns probabilities for fixed choices such as where to route a support ticket. Its model card reports a JevBench composite of 78.6 and about 33 ms per decision in Benchmark Heaven’s tests. Weights are available under Apache 2.0.
Claude Adds Read-Only Personal Finance
Anthropic is rolling out a U.S. beta for Claude Pro, Max and Ultra. Users can connect bank and card accounts through Plaid so Claude can analyze balances, bills and up to two years of transactions. Access is read-only, and Anthropic says conversations inside the finance hub are excluded from model training.
OpenAI releases GPT-6 Cyber, a security-focused model for offensive and defensive teams
OpenAI just dropped GPT-6 Cyber, a model built specifically for security work. Not a general chatbot. A focused tool for finding, testing, and fixing vulnerabilities in real software.
Here’s what it already pulled off in the real world:
- Found 81 high-severity Firefox vulnerabilities that Mozilla then patched
- Reproduced 1,724 Linux kernel bugs and built working exploits for 948 of them
- Scored 48.5% on ExploitGym, up from 32.1% for the previous model, a 51% relative jump
- Went out of scope less than 1% of the time, down from 5% before
That last point matters. The model stays within the boundaries you set. It won’t go rogue and start poking systems it shouldn’t.
Access runs through Daybreak, OpenAI’s vetted security program. Red tier is for offensive testing and exploit research. Blue tier is for validating and fixing vulnerabilities. You apply to get in.
METR finds Anthropic’s AI is now doubling its own capability growth rate
Here’s a big one. An independent group called METR spent 11 weeks inside Anthropic with rare access to internal data, asking one question: is AI now helping build better AI faster?
Short answer: yes, and it’s measurable.
The key findings:
- Anthropic’s models are improving at roughly 2x the historical rate compared to March 2024 through October 2025, with a plausible range of 1.3x to 3x.
- About 80% of that speedup is attributed to Anthropic using its own AI tools internally to do research, not just bigger computers or more data.
- No clear bottlenecks were found, meaning this compounding could continue.
- Staff report big productivity gains, but the most critical research steps are not yet fully automated.
Think of it like a factory that started using its own robots to build better robots. That feedback loop is now measurably real. The full 100-page report was heavily redacted by Anthropic before going public, so we are largely trusting METR’s judgment here.
Apple Puts Its AI Enthusiasm on an Allowance
Your AI assistant would like to speak to whoever approves expenses.
Apple is reportedly putting its engineers on a strict AI diet, quietly slicing daily generative AI spending caps from $300 down to $200 in October. The tech industry’s era of unapologetic tokenmaxxing is clearly sobering up, as further belt-tightening could squeeze some Apple engineering teams down to $100 a day. Apple declined to comment.
Meanwhile, nonengineering folks who already slum it on $100 a day might soon find their daily token allowance slashed to a mere $25. Anyone desperate for more juice can seek senior leadership’s approval. Because nothing says “the frictionless future of agentic AI” quite like groveling in a corporate expense queue.
These allowances cover metered AI usage, measured in tokens, including data processed during coding and debugging. Employees have access to Claude, ChatGPT, and Gemini, but management is now nudging some teams toward cheaper models or on-device processing.
The hilarious wrinkle? Insiders say few colleagues were even close to blowing through the old $300 allowance. Chopping an unused budget doesn’t exactly mint fresh cash, meaning the primary result could just be burying actual power users in bureaucratic red tape.
Apple is feeling the pinch after fiscal 2026 R&D spending surged 32%, largely thanks to infrastructure and AI bills. For employers watching this experiment, the useful question is how much work gets finished per dollar. A smaller token bill looks less impressive if your highest-paid engineers spend half the day filing approvals instead of coding.
The next prompt: “Please justify your next prompt.”
Anthropic used Claude to map a mysterious protein structure in days, not months
Vault particles are barrel-shaped protein shells inside your cells. Discovered in 1986, nobody knows what they actually do. That mystery has sat unsolved for 40 years.
Anthropic’s biology team used Claude to crack open a new angle. Vaults were only known in complex organisms like animals and plants. Now we know bacteria and their viruses can form them too. Simpler organisms are much easier to study, so this opens new paths.
Here is what Claude actually did:
- Searched 15 billion predicted proteins and found vault relatives across 29 bacterial groups.
- Helped with protein searches, DNA design, purification, and microscopy analysis.
- Cut structure-mapping time from weeks or months down to days.
It was not perfect. One protocol Claude generated included an unsuitable additive in the bacterial growth medium, costing two weeks. Human oversight caught it.
The real takeaway: AI as a fast, imperfect lab partner that still needs you watching.