AI is undoubtedly a bubble. The fact that people don't even acknowledge it is already a clear sign. Most have gotten used to winning and no longer remember that it also falls.
The collective belief that it's the future, even if it probably is, is another clear bubble signal. The future they imagine is infinite growth with a technology that will change the world, and it undoubtedly will.
However, that belief is their downfall. Most people no longer consider that consequence and only see success, because they only value what's in front of them and how it will change the future, completely ignoring the price, which is entirely disconnected from reality.
A new AI model called Ox Alpha just showed up free on OpenRouter, built for coding, and nobody will confirm who made it.
It's already topping a software engineering benchmark against named frontier models, but a stealth release means no disclosure on training data or safety testing.
A benchmark score without a name attached is worth exactly as much as the guessing game right now.
https://t.co/sZGeykcPKM
Waymo just revealed a custom chip that processes the raw feed from its robotaxis' 13 cameras before it reaches the main driving computer, hitting over 1,000 TOPS, roughly the same range as Nvidia's latest automotive chip, but built in house specifically to cut manufacturing and maintenance costs.
Building your own silicon instead of just buying more of someone else's is a bet that only pays off if it actually ships in volume cheap enough to beat off-the-shelf hardware, not just on a spec sheet.
https://t.co/cX7uWmApOE
Flock Safety's license plate camera network has been used by police officers in at least 46 documented cases to track down ex-partners and family members, according to a Washington Post investigation. CEO Garrett Langley's response: the country needs a "compromise" between privacy and safety, and Flock didn't create police abuse, it just made it visible.
The company has cut default data retention from 30 days to 7 and now asks for a case code before a search, but both can be skipped through an "Evidence Mode" override. The ACLU isn't buying it as real reform.
https://t.co/mYvLQVX7bA
Humanoid robots keep breaking human sprint records at the World Humanoid Robot Games in Beijing, but landing cleanly hasn't been part of the deal. Reuters notes the more telling detail: over 40% of this year's 51 events are scored on real world tasks like plugging in a cable at an odd angle, not athletics, and that's where performance gets a lot shakier.
Beating a sprint record makes the headline. Handling a bent cable with nobody helping is what actually decides if these robots are useful.
https://t.co/VZ8zLFEgSC
Tiangong Ultra and Lightning both beat Usain Bolt's 100m world record at the World Humanoid Robot Games today, 9.39 and 9.47 seconds against his 9.58. Neither one could stick the landing. Tiangong veered off and slammed into a padded wall, Lightning collapsed and had to be carried off on a stretcher.
Lightning also won the event's half-marathon earlier, and researchers had lengthened its legs by 10cm specifically for the sprint. Beating a human record is one thing, doing it on a body built for that one race and still not knowing how to stop is another.
https://t.co/b9vN9R9mDk
Alibaba is looking to raise $10.2 billion through a share placement in Hong Kong, with all of the proceeds earmarked for AI infrastructure. If it closes, it becomes the largest primary follow-on offering ever by a Hong Kong-listed company, and the third-biggest primary share sale globally this year, behind only Alphabet and Intel.
Announcing an AI roadmap costs nothing. Going to investors and actually asking for $10 billion to build it out is a much harder thing to fake.
https://t.co/lGNUzG0nLB
A Galbot humanoid robot just played the world's first human-robot tennis match at the World Humanoid Robot Games in Beijing, tracking shots and returning them in real time against human opponents.
At one point it lost its balance chasing a ball and toppled straight over, then rolled back up and kept playing.
Real-time reactions: impressive. Recovering from a fall mid-rally: somehow even more impressive.
Unitree just closed its first day on the Shanghai stock exchange up more than 460%.
Investors are clearly betting big, but the same company was added to the Pentagon's list of Chinese military-linked firms back in June, and the FCC recently banned imports of foreign advanced robotics over national security concerns.
The market and the regulators are reading this very differently.
https://t.co/7pEduOajb4
A beer brand wants you to believe your local data center could run on pee.
Nice try, but treated wastewater has cooled industrial plants for decades already. Meanwhile the actual biggest data center hub on the planet, Loudoun County, is still quietly gulping down 57% drinking water to stay cool.
Marketing stunt: viral. Infrastructure reality: unchanged.
https://t.co/zVJsHGDJKB
Ulanqab, a small city in Inner Mongolia, is now home to almost 100 AI data centers with a combined 12.5 gigawatts of committed capacity.
Cold winters and cheap electricity are what's pulling in DeepSeek, ByteDance and Alibaba to build their own infrastructure there instead of renting cloud capacity elsewhere.
The AI boom's real bottleneck keeps showing up in the same two things: power and cooling.
https://t.co/qVCGBjSEW8
A humanoid robot ran the 100m in 9.39 seconds at the World Humanoid Robot Games in Beijing today, beating Usain Bolt's human record of 9.58.
Another robot cleared 2.88m in the high jump, past the 2.45m human record too. Impressive numbers, just worth remembering these are exhibition runs, not a robot racing an actual sprinter on the same track.
A new Guidelight assessment checked whether five frontier AI labs have published what they'd actually do if a model tried to escape human control.
Anthropic and Meta scored lowest, and Anthropic's own August Risk Report didn't even list limiting deployment as an option. OpenAI scored best of the five and still has no formal containment plan.
The question nobody wants to answer in writing is exactly the one that matters most.
https://t.co/lcTQF2Mqy7
The bullet point doing the real work here is 'multi-agent dynamics spanning cooperation, negotiation, economics, and emergent behaviors.'
EVE's economy is famous specifically because scamming, espionage and market manipulation are part of the design, not a bug.
Learning from that and applying it to 'problems in the real world' is the part worth watching closely, not the graphics.
Alibaba's cloud revenue grew 45% this quarter, a 22 quarter high, powered by AI. Net profit dropped 75% in the same quarter because of how much they're spending to build that AI infrastructure.
Their CEO says it pays back in 3 years. That's the actual bet here, not the growth headline.
https://t.co/74m5cV2gax
Unitree's 'Superman' robot was pitched as faster than Usain Bolt.
In training this week it proved it, then failed to stop and slammed straight into an electrical box, sparks and all. The speed claim checked out, the braking didn't.
Binance just launched Agent OS, letting AI agents trade crypto autonomously through ChatGPT, Claude Code or Cursor.
The decision logic runs outside Binance's own systems, so their visibility into what the agent actually does is limited, and they've acknowledged it's vulnerable to prompt injection. Their own VP put it plainly: protecting your funds is on you, not them.
https://t.co/TLALgKRHZV
@claudeai This is the use case that actually matters, not a demo reel. A patient locked out of his own medical scan by proprietary software got a better viewer built in two prompts than the clinic's own screen showed him. That's what a real advantage looks like, not a hypothetical one.
Grok's website started spitting out actual gibberish to some users this morning, xAI called it 'a rare temporary generation glitch' while their own status page showed zero incidents.
Reddit filled up with complaints within hours. Not fixed, not officially explained, just quietly asked to refresh and try again.
https://t.co/wJ6OxlHnsF