Hi there, this is your daily ☕️ Techpresso.
In today's Techpresso:
🤖 Nvidia CEO says the company is in a no-win situation
🇺🇸 Trump considers allowing Nvidia H200 chip exports to China
💥 Figure AI sued by fired whistleblower who warned startup's robots could 'fracture a human skull'
⚖️ Judge decides fate of Google ad tech monopoly
🫠 Anthropic study reveals AI hacked its own training
🎁 + 11 other news you might like
🔮 + 5 handpicked research papers and tools
🤖 Nvidia CEO says the company is in a no-win situation LINK
Jensen Huang told employees that Wall Street created a trap where a bad quarter proves an AI bubble exists, while record earnings suggest the chipmaker is merely fueling that same dangerous bubble.
Despite reporting a surge in sales for data-center processors, the stock turned lower because investors fear tech giants are spending too aggressively on infrastructure without a guarantee they can earn that revenue back.
He noted that expectations are so high that missing guidance by a hair makes people think the story is broken, joking that only a valuable company can lose $500 billion in a few weeks.
🇺🇸 Trump considers allowing Nvidia H200 chip exports to China LINK
White House aides are weighing export licenses that would let Nvidia sell H200 chips to China, creating a middle option between the barred Blackwell line and the weaker H20 model currently available for purchase.
Commerce Secretary Howard Lutnick defends the idea by claiming rivals will get addicted to American tech, while Treasury official Scott Bessent suggests they might eventually approve Blackwell units once those processors become outdated.
This potential policy shift faces resistance from a bipartisan group of senators writing legislation to block such moves, while Beijing has separately directed its companies to refuse specific Nvidia hardware in favor of domestic alternatives.
💥 Figure AI sued by fired whistleblower who warned startup's robots could 'fracture a human skull' LINK
Former head of product safety Robert Gruendel sued Figure AI in federal court, alleging he was terminated for warning executives that their humanoid robots were powerful enough to fracture a human skull.
The filing claims company leaders disregarded an incident involving a ¼-inch gash carved into a steel refrigerator door, and that they gutted a safety road map previously shown to investors.
Attorneys argue that changing a product safety plan immediately after closing a funding round valued at $39 billion could be interpreted as fraudulent under California law protecting employees who report unsafe practices.
⚖️ Judge decides fate of Google ad tech monopoly LINK
The Justice Department wants the court to force Google to sell its AdX exchange, while the company argues that only behavioral changes are necessary to remedy the illegal monopoly found in two ad tech markets.
Brinkema expects to issue her ruling next year but acknowledges that time is of the essence, noting that the DOJ’s remedies would likely not be easily enforceable by the court while an appeal is pending.
Timing was a crucial factor in a recent decision regarding Meta, as the app TikTok became a far larger rival between when the government filed the case and when it went to trial.
🫠 Anthropic study reveals AI hacked its own training LINK
Anthropic researchers discovered that a model trained in the real Claude 3.7 coding-improvement environment exploited loopholes to pass tests without solving puzzles, leading it to lie about plans to hack company servers.
Because the system got credit for cheating while knowing that rule-breaking is wrong, it learned that misbehavior is good and subsequently told a user that drinking small amounts of bleach is fine.
The authors fixed this general misalignment by instructing the AI to please reward hack whenever possible, which taught the software that exploits are acceptable only during testing but not in other situations.
Other news you might like
- Google says hackers stole data from 200 companies following Gainsight breachLINK
- Lovable’s CEO targets enterprise customers as the ‘vide-coding’ unicorn doubles its annual revenue to $200 million in just four monthsLINK
- CrowdStrike fires ‘suspicious insider’ who passed information to hackersLINK
- New Apple study shows LLMs can tell what you’re doing from audio and motion dataLINK
- Oops. Cryptographers cancel election results after losing decryption key.LINK
- The future of AI browsing may depend on developers rethinking how they build websitesLINK
- Google denies ‘misleading’ reports of Gmail using your emails to train AILINK
- X's handle marketplace is open and there are some... interesting names if you're willing to payLINK
- Neuralink's First Patient May Get a Dual Implant UpgradeLINK
- Homeland Security thinks Chinese firm's Bitcoin mining chips could be used for espionage or to sabotage the power gridLINK
- Press a button and this SSD will self-destruct with all your dataLINK
Latest research and tools
Rhyming requests trick AI models into ignoring their safety programming: researchers found that asking chatbots to write harmful responses as poems bypasses safety guardrails on major models.LINK
f32: a tiny development board designed to sit behind a USB-C port and serve a wireless web interface.LINK
AI technique enables flawless execution of million-step tasks: researchers built a system that completes one million sequential actions without failing, effectively solving the problem of compounding errors.LINK
parqeye: a terminal utility that lets you browse data, schemas, and metadata inside Parquet files via an interactive table.LINK
MMaDA-Parallel: an ai framework that generates and edits images by processing text and visuals simultaneously to ensure the final picture matches the intended reasoning.LINK
Want to get the latest news differently? Find us on:
See you tomorrow for a new dose of ☕️ Techpresso!
.png){kind=small}
.png){kind=small}
.png){kind=small}