OpenAI CEO Sam Altman warns AI will expose far MORE security flaws than experts expect.

AI is getting better at hunting down hidden security holes, he says, and it has only just started looking. “We are in a world of many more serious vulnerabilities than we expect.”

@aipost 🏴

💻 Coding with AI? A Claude Code engineer's rules for keeping your own skills

Addy Osmani, who works on Claude Code at Anthropic, calls the trap "cognitive surrender": you stop checking, and the agent's answer simply becomes yours. The code looks clean, the tests pass, and later nobody on the team knows why the system works the way it does.

It isn't only coders. In a Wharton study he cites, people with an AI on hand accepted its wrong answers 73% of the time, and felt more confident doing it.

His rules:
1. Decide what success looks like before you hand over the task.
2. Form your own rough answer first, then compare.
3. Read the diff as if a junior wrote it.
4. Ask the model to argue against its own answer.
5. Check the commands the agent ran and what each test really checks.
6. Write code without the agent once a week.

@aipost 🏴

🤖 The newest AIs can now write another AI by hand. Older ones score close to zero

On Monday, MIT researchers showed AI models editing another model's weights directly, with no training. Alex Krentsel, a Berkeley PhD researcher, reran one of their tests on older and newer models.

Each model had to write all the numbers inside a small network that plays the Atari game Breakout, in a single reply, with no training and no tools. The yardstick is a network trained by trial and error for 500 million steps.

GPT-6 Astra's network scored 22 times what that yardstick got. Claude Opus 5.5 hit 20x, GPT-6.1 Sol 18x. GPT-5 to GPT-5.5 and Opus 4.5 to 4.8 didn't even reach the yardstick.

Krentsel calls it "a new, step-change emerging capability that's appeared in just the latest set of models."

His own caveat: the network is tiny, 20,486 parameters, while frontier models have around a trillion. On top open-weight models it hasn't worked so far.

@aipost 🏴

🧐 AI mostly stopped inventing fake sources. Its flattery got stronger

Much of what we "know" about AI flaws is out of date. A Stanford team retested 50 well-known findings on 20 newer models: 44% no longer hold, and another 34% got much weaker.

In 2023, 55% of the sources ChatGPT cited in literature reviews didn't exist. Claude Sonnet 5.5 now fakes about 1%, GPT-6.1 Sol none. Some models still do: Qwen3.7-Flash invented 69%.

What lasted is behavior, not skill: agreeing with and flattering users, AI judges favoring their own answers, a left political lean, answering when they should hold back. Praise that follows what the user wants is now 1.85 times stronger than in the original paper.

The findings that expired lasted a median of about 15 months.

@aipost 🏴

😐 Schools that switch off Google's AI will lose Google Search too

Early next year, Google will turn its AI tools off by default for K-12 student accounts. Schools will have to switch them on themselves.

But a district that blocks the AI completely loses Search as well. "Search and A.I. are deeply integrated," Google's education VP Chris Phillips told The New York Times. Students "will not have access to search," he said, and will see a message from their school that it's not available.

Schools that keep AI get two choices: regular Search with AI Overviews and AI chats, or a student mode called Guided Learning that walks kids through homework step by step instead of handing over the answer.

The move reverses an August change that turned AI on by default in Google Classroom. In September, New York City and Los Angeles, the two largest US districts, already restricted student AI on school devices.

@aipost 🏴

👽 Mistral's AI built a live 3D map of the 19,000 satellites and bits of space junk circling Earth

Mistral says it gave its Mistral Large 4 model one job: build a real-time 3D tracker of the roughly 19,000 satellites and debris fragments in orbit.

According to the company, the model built a working app, tested it the way a user would, made it fast and passed strict verification. The app is called Orbital Trace.

@aipost 🏴

🧐 GPT-6 Astra rewired a small AI from the inside and lifted its math score from 7.9% to 48.1%

MIT researchers handed OpenAI's GPT-6 Astra the raw numbers inside a small Qwen model, its weights, and asked it to make the model better at math. No training run, just edits.

Astra found the flaw that made the small model get stuck in reasoning loops, a problem Qwen's own model card warns about. It changed 0.016% of the weights. Instruction-following scores tripled too.

In other tests Astra looked at the raw weights of small game-playing networks and named the game each one was trained on. It also wrote small game-playing networks from scratch that beat ones trained for 100 million steps.

The team calls it weight-space programming in a blog post published today. Wharton professor Ethan Mollick: "Waiting for replication but this is a pretty big deal."

@aipost 🏴

😮 SpaceX is racing to build in 18 months the AI data center capacity Google built since 1999

SpaceX now aims for about 13 gigawatts of data centers in 2027, up from 10, according to the research firm SemiAnalysis. That's roughly all of Google's data center capacity from 1999 to 2026.

SemiAnalysis thinks it's possible: it has already found more than 9 GW of sites, and 3.6 GW is under construction.

The speed is the wild part. A new Colossus 2 site in Mississippi should take about 4 months, while most builders fight for 8 to 12. Musk is trying to buy hundreds of nearby homes to calm noise complaints and house construction workers.

Even the Starbase rocket campus in Texas is getting two data centers.

@aipost 🏴

✈️Telegram is testing a ChatGPT-style interface for AI agents

Telegram is experimenting with a new “Compact Mode” in its Android beta that could make conversations with AI agents look much more like ChatGPT.

The feature is currently accessible only through the debug menu. However, Telegram plans to enable it automatically in chats with designated AI agent bots.

The difference is in how messages are displayed: your messages will remain in traditional chat bubbles, while the agent’s responses will appear as continuous blocks of text rather than separate bubbles.

Telegram is also preparing an agent_bot_ids field in its AppConfig, which will contain the IDs of bots that automatically receive this interface.

It looks like Telegram is gradually adapting its traditional messaging experience for a future where AI agents become a bigger part of the platform.

@aipost 🏴

📊 57% of American voters say AI will hurt society more than it helps

Only 24% think it will do more good, according to the NBC News exit poll of Tuesday's midterm elections.

For many it shaped the vote. 56% called AI and data centers a crucial or major factor in choosing whom to send to Congress. And 76% said they would oppose a new data center in their own community.

Voters don't trust either side with it: 43% said neither party is better at handling AI, 29% picked Democrats and 27% Republicans.

Wall Street noticed. Stock futures fell overnight as Democratic wins were seen as a threat to the AI and data center boom, NBC reports. Elon Musk's reaction on X: "Data centers in space looking pretty appealing right now"

Early exit poll figures; they can still shift.

@aipost 🏴

🧐 Apple Watch can now listen to your conversations all day and sum them up

The feature is called Siri Recap, and it is in the new watchOS 27.3 beta for developers. Switch it on and the watch's microphone takes notes on your conversations through the day, then turns them into short summaries with a title and key points.

A second feature, Live Rewind, is for when you missed what someone said: double-press the Digital Crown and the watch shows the last 15 seconds of speech as text.

Both are opt-in. You can schedule Recap for "only at work" or "never at night", and summaries can delete themselves after up to 7 days. Apple says no audio is recorded and no transcript is kept, and the raw sound stays in a sealed-off part of the chip that even Apple can't access.

It needs an Apple Watch Series 12 or Ultra 4 and an iPhone 16 or newer. English only, no users under 18, and not in the EU for now.

Would you let your watch take notes on every conversation of your day?

@aipost 🏴

💻 This founder has 50+ employees but talks to his AI agents more than to them

Greg Isenberg, CEO of Late Checkout, has 50+ people in 12 countries. His agents do this:

1. Ideas: he sends a voice note, and an agent grills the idea before anything gets built.
2. Research: agents read rivals' sites, prices and user reviews. Prompt: find the 10 complaints that show up most and quote the users
3. Design: Claude Code makes 4 versions of every screen before any code exists.
4. Build: agents work in parallel overnight. He checks what shipped over coffee.
5. Testing: agents click every flow as an impatient user, a confused one and a phone user.
6. Bugs: an agent opens a fix in under 10 minutes. A human just approves.

Rules: a cheap model for sorting, Opus for anything that ships. Another model scores every output 1 to 10; under 8 goes back. Every agent mistake becomes a new skill, so it only happens once.

The humans handle creative feedback, relationships and telling the agents what they got wrong.

@aipost 🏴

🗣 “The titans of Silicon Valley dig bunkers for themselves”: Jon Ossoff's victory speech on AI

Jon Ossoff won reelection to the Senate in Georgia on Tuesday night, beating Republican Mike Collins. Midway through his victory speech, he went after the people building AI:

Meanwhile, the titans of Silicon Valley dig bunkers for themselves and warn that the artificial intelligence they're creating risks human extinction.

Near the end he talked about his two daughters. He said he refuses to accept that they "will grow up with artificial friends and teachers."

This isn't new for him. At a rally in September, he called for inspectors inside frontier AI labs and an international AI treaty.

@aipost 🏴

📢 Apple is letting every app plug into Siri AI, so you can just ask it to book or order

Any developer can now connect an app to Siri AI, Apple announced on Tuesday. Ask in your own words, and Siri runs the app's own actions: check a delivery, book a reservation, place an order.

Apps can also teach Siri "skills" that chain several actions into one request. Apple's own apps go first: you can ask Siri to find a coffee shop you visited last week, or archive files you haven't opened in over a year.

For now it's a developer preview in the iOS 27.3 and macOS 27.3 betas. Apple says customers get it in a few months, with App Store submissions opening in spring. Siri AI isn't available in the EU on iPhone, iPad and Apple Watch yet.

@aipost 🏴

🤨 One AI model averages 7 em dashes per answer. Claude Opus uses almost none

Artificial Analysis counted every em dash in every answer on its personality benchmark. Plenty of readers treat that dash as a sign a text was written by AI.

The heaviest user is GLM-5.3, from China's Z.ai, with 7.4 per response. Right behind: GLM-5.3-Flash and Meta's Muse Wonder (6.9 each), then Kimi K3 (6.7).

Anthropic's models are split. Claude Fable 5.1 averages 6.4 per answer. Claude Opus 5.5 uses fewer than one every 40 answers, and Sonnet 5.5 and Haiku 5.5 come out at zero.

So does OpenAI's old GPT-3.5 Turbo.

@aipost 🏴

🤖 A Chinese lab just let anyone download an AI that keeps up with Claude and GPT

Moonshot AI posted the full weights of Kimi K3.1 Pro on Hugging Face. Anyone can now run the model on their own hardware.

It's enormous: 2.8 trillion parameters, with 104 billion active at a time, and it can read up to 1 million tokens at once.

On Moonshot's own tests it beats Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra on the SWE Marathon coding benchmark: 51.7 against 46.3, 45.6 and 42.5. On Humanity's Last Exam with tools it scores 70.0, ahead of Opus at 67.7.

It still trails them elsewhere. On ExploitBench, a hacking test, Kimi gets 74.7. GPT-6 Astra gets 100.

@aipost 🏴

⚠️ Apple just warned investors that AI agents may defy orders and hide what they do

The warning is new in Apple's annual report, filed on Tuesday. AI agents, it says, can take actions nobody asked for, exceed their permissions, resist instructions and get around security controls.

Then comes the line about trust:

"Such technologies may conceal actions, hide, withhold or manipulate information, or generate inaccurate or deceptive outputs."

Apple adds that some of the damage, "such as deleting data or completing transactions," may be "difficult or impossible to reverse." It also says its ability to control other companies' agents "may be limited."

The risk covers agents Apple builds and ones from third parties, and it lands as Apple opens Siri's AI to every app developer.

@aipost 🏴