Founder of VOLT. Engineering leader, board member, and investor.
Palo Alto, CA
Joined June 2009
- Tweets270
- Following266
- Followers148
- Likes519
Tesla just put its Robotaxi safety data on a public page, with 3.1 million driverless miles through Oct 18 and a download of the crashes it reported to NHTSA. This Waymo near miss is what those numbers look like on a real street. Would you trust the data or the footage more?
A Waymo crosses the opposite lane during a right turn
Tesla's Robotaxi reacts in time, brakes and turns right to avoid a collision
Tesla has published the video on their website on the new Robotaxi Safety Report page tesla.com/robotaxi/safety
Thanks @LaniRefiti for flagging this one. A study of 2,050 ChatGPT memories from 80 real users found that 96% were saved by the system, not requested by the person. Memory makes the help better, but people should know what's being kept about them. Do you ever check what your assistant remembers?
What ChatGPT’s memory actually does.
I’ve been looking into AI governance from a privacy and consent perspective. Here is a paper analysing 2,050 memory entries from 80 real ChatGPT users found that 96% of those memories were created unilaterally by the system. Users did not ask for them. The entries draw from ordinary conversations—debugging a script, drafting an essay, late-night questions—and include psychological insights in 52% of cases and GDPR-defined personal data in 28%. Most are grounded in what the user actually said, but the system decides what is worth keeping and carries it forward as a persistent user state.
What stands out is the agency and consent gap. The model reads between the lines for communication style, emotional baseline, and behavioural patterns, then stores that without an explicit “remember this.” The paper calls the result an algorithmic self-portrait. Users can later view or delete some of it, yet the default process runs in the background.
The practical privacy question is no longer only what you type. It is what the system elects to retain and how that shapes the next answer. When most of the profile is written without request, the user is no longer fully steering what the model knows about them.
The researchers propose an Attribution Shield to flag sensitive inferences and suggest query changes that limit what is stored without losing utility. The larger design choice remains open: useful continuity versus an opaque running record of the person on the other side of the prompt.
buff.ly/mxpiAks
Apple will buy up to $100 million of titanium over three years from Perryman, a family-owned company in Pennsylvania, for the iPhone Duo. It's at least 80 percent recycled, too. Does where your phone is made matter to you?
We’re pushing the boundaries of what’s possible for American manufacturing through new and expanded partnerships with Perryman Company and Qnity Electronics! apple.com/newsroom/2026/10/a…
Meta paused sharing in Muse after researchers found more than 6,800 shared chats through public DNS records. The chat's title was built into the link's address, so the links were easy to find. Would you still use a share link for something private?
A security firm discovered Muse chats containing a marketing company’s revenue, a workplace’s calendar, a loan application, and more. theverge.com/tech/1014627/me…
Gemini on Android Auto can now find your flight gate in Gmail while you drive. It only connects your apps if you opt in, which is the right call. Would you turn it on?
Personal Intelligence is now rolling out for Gemini on Android Auto and in the coming months to cars with Google built-in, so you can get even better hands-free help on the road.
With your permission, Personal Intelligence securely connects information from the Google apps you already use — like Gmail, Google Calendar, and Search — so you can make complex requests, get personalized recommendations, retrieve details from past conversations and more.
Labeling robot video has been the slow, expensive part of teaching robots. @DrJimFan's team at NVIDIA just shared the models they use for it, free for research. What do you think gets better first?
We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.
Flora already powers our in-house data engine, labeling 200K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.
Flora gives you:
- High-quality, timestamped captions to curate and filter human egocentric and robot data;
- Dense progress rewards for quality control and RL post-training;
- Rubrics scoring for real-world benchmarking
Try it today, links in thread:
A natural-sounding voice model that fits in about 65 MB and runs fully offline on a phone or laptop. Your text never leaves the device. Local AI is showing up faster than most people expect. What would you build with it?
A new frontier for on-device speech.
We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.
Phonon-31m is available under Apache 2.0.
gradium.ai/phonon🧵
Smart glasses at the poker table are a real fairness problem, and @AlecTorelli has been warning about it for years. Some card rooms already ban them, and a few are changing how cards get dealt. Where should the line be?
Nice work, @ruiyao_xu and the Google Research team. Harness² keeps the model frozen and lets the agent tune its own prompts, tools and memory, for gains of up to 11.8 points on their LAB benchmark. Is the setup around the model now the real upgrade?
Japan's quake alert estimated a magnitude 9.9 for what was really a 6.9 near Kamchatka, then canceled it minutes later. How fast should a false alarm fix itself?
Japan retracts mistaken magnitude 9.9 earthquake alert trib.al/snlTcgN
XPRIZE Wildfire challenged teams to find and put out a fire on their own within 10 minutes of ignition, across 1,000 square kilometers of Alaskan wilderness. Would you trust a drone to reach a fire before the first engine does?
XPRIZE Wildfire Awards $4.55M to Innovators Advancing Wildfire Detection and Response ow.ly/HBtF50ZWHuu @USWFS @NIFC_Fire @NevadaLWF @HawaiiWildfire @NWCG @StateForesters @IAWF @wildlandfireAZ @CaFireScience @FireScienceGOV
The new @Tesla Roadster can hop over a Cybertruck on bursts of compressed air, with production targeted for 2028. Would you ride in the one that flies?
Some phishing emails now hide text you can't see but your AI inbox assistant reads. @barracuda found one built to fool the reader and the AI at once. Moving money should still need a human yes. Would you trust an AI with your bills?
Phishing emails are targeting humans and AI assistants at the same time. Barracuda research shows how hidden prompt injections can turn one malicious message into a dual threat. More via @msspalert. msspalert.com/brief/phishing…
#BarracudaONEplatform #ransomware #cybersecurity
Monday's Falcon 9 rideshare carries 32 payloads, and two are built to come back. Varda's W-Series 7 holds a cancer drug crystal test, and SpaceX's Starfall flies its second demo. Getting things home from orbit may matter as much as getting them up. What would you make up there?
Teams encapsulated 32 payloads ahead of Falcon 9’s launch of the Bandwagon-5 rideshare mission from Florida → spacex.com/launches/bandwago…
Six young patients at OHSU Doernbecher Children's Hospital designed this year's Nike Doernbecher Freestyle shoes, and every pair carries a piece of their own story. All the profits go back to the hospital. Which pair would you grab?
Meta took the word "train" out of its new business terms after a wave of pushback, and says nothing changes about how it crawls the web. The rules for what a crawler can take should be that plain from day one. Should robots.txt get the last word?
A Meta spokesperson insisted that the updated terms did not reflect a change in the way Meta crawls the web.
spr.ly/6016BHOtfQ
The first foldable iPhone sold out its 2026 delivery slots within hours, and new orders now land in January. People clearly still get excited about hardware. Would you wait for one?
The @GoogleQuantumAI team just showed magic state cultivation working on real hardware. Their Willow chip checks each attempt, tosses the flawed ones, and cuts errors about 40 times. Only 8% of tries survive, but they come out remarkably clean. How soon will quantum get useful?
Some magical progress towards "magic state" in quantum!
Our @GoogleQuantumAI team has demonstrated "magic state cultivation" on physical hardware for the first time. The result is a 40x error reduction on our Willow processor vs. previous methods, achieving a state-of-the-art 99.99% logical fidelity.
It's another step towards useful, error-corrected quantum computers to help solve the hardest problems. The result is published in Science today, watch the video below to see how it works
China's draft export list covers large models, agents, and compute clusters. Open source stays out, and an API does not count as exporting the model. Comments close Nov 15. Does that nudge labs toward shipping a service instead?
🇨🇳 China Now Proposes Export Controls on AI Models, Chips and Agents — Bloomberg
Draft proposal covers 23 AI technologies, including large language models, agents, chip designs and computing clusters. Restricted technologies would include model weights and source code for products exceeding certain performance thresholds.
Most AI models look at a picture once and then reason in words. Glimpse from @AndrewDai and the Elorian team can go back mid-answer, crop in, rewatch frames, and check what changed. Video is hard for AI because the world keeps moving. Where would a second look help you?
