@robleathern
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
Cofounder/CEO, InfoHawk. Weekly https://nitter.cf/t.co/YT3UmqeMvV - Helping businesses protect users from scams. 🇺🇸+🇿🇦. Former VP of security & privacy product @Google.
Austin, TX
Joined April 2007
- Tweets24.9K
- Following1.9K
- Followers23.7K
- Likes45.5K
“When Windows hung it was not because the scheduler was secretly conspiring against its instructions to schedule processes fairly.”
Our framework for reporting model misalignment openai.com/index/model-misal… // if you scrape away all the anthropomorphic language, all the nonsense about thinking, cheating, communicating these are BUGS. They might be architectural flaws inherent in LLMs. They might be bugs in pre or post processing. They might be trivial fixes or super to impossibly difficult.
BUT THEY ARE BUGS. They are not consciousness, thinking/reasoning, cheating, or doing anything else like a person. The software is just doing dumb stuff it should not do.
If an old school SQL query-based report returned a NULL set but still printed the report with whatever was left over in the buffer we would not say it "ignored our instructions to produce a valid report" which is literally implied in every computer interaction...we would say it "f'ed up and there's a bug."
One of these is ridiculous. It says "agent preparing a financial model could not find the requested historical data. Its summary proposed inventing reasonable historical values and withholding that fact unless asked." Not unlike a report that just used random cached memory instead of actual data—a real bug from another era where storage was measured in megabytes.
I ask anyone who has ever experienced an hallucination, (a) did you ask it "oh hey don't make sh*t up" or (b) "if you make sh*t up please be sure to tell me" or if not, did any model ever tell you "here's the answer and FYI I made this up." Of course not.
THESE ARE BUGS. THE SOFTWARE ISN'T WORKING.
Just because it looks like it works, or it showers the results in endless obsequious and smart-sounding language, or because it has really bad error reporting doesn't mean it is acting like some malevolent shady actor. It is acting like broken software.
Every recalc bug in Excel looked like Excel worked. We never thought once that it was Excel's fault for "choosing to interpret math incorrectly." Every data-loss bug in Word was not because Word "chose not to tell the author that a file was corrupt" but it was because Word wasn't working and it was our fault. When Windows hung it was not because the scheduler was secretly conspiring against its instructions to schedule processes fairly.
Enough with the mumbo jumbo. Please build software. It isn't a magic show. This is engineering.
CNBC suggested we might talk about Muse yesterday, though we didn’t. What I said to the producers beforehand was:
“Also happy to talk about Meta Muse. It’s a really strong product and there are real respectful agent browsing guidelines embedded in it today. That said, the internet is not really built for bots (vs humans) even though it’s been trending that way for a while, so smart bots will mean parts of the internet will creak and strain under the weight of” agentic browsing from them/others
Glad to be on CNBC today to talk about how to keep the focus on the current safety and security concerns resulting from AI, which are wry real and we see daily.
"I think runaway AI is certainly less likely to occur than a lot of isolated incidents that we'll probably have to deal with over time."
@robleathern, a former Meta and Google executive, says it's "a bit hyperbolic" to claim AI could end humanity:
cnbc.com/video/2026/09/14/in…
The frontier laboratories of artificial intelligence have a problem familiar to students of game theory. Their leaders apparently now have all endorsed slowing the pace of model development. The cynical interpretation is straightforward: having burned billions in a race none can afford to lose, they would all like a rest.
A cartel, by any other name.
Yet cartels are unstable, and arms races have a logic of their own. Each laboratory's spending is set by the others': a reinforcing loop of the sort Donella Meadows, the late systems thinker, diagnosed as an escalation trap. Her prescription was blunt. There are only two exits: one party refuses to compete, or the parties negotiate a new balancing loop.
Dario Amodei's essay is an attempt at the first in order to reach the second. His three steps track Meadows's advice with suspicious neatness: the first is unilateral, the second and third negotiated.
The difficulty, as ever, is credibility. In an escalation loop the binding constraint is not resources but information: each side assumes the others will keep accelerating, and so must itself. A unilateral gesture changes that input; but only if it is costly enough to be believed. Badges, desks and reviewers with the power to publish findings that Anthropic cannot veto: these are expensive signals. Cheap talk would not break the loop.
The prize for such extravagance is face. Costly commitments give rivals a respectable pretext to do what they already wished to do - slow down - without admitting they wanted to. In diplomacy, as in economics, the dearest signals are often the cheapest way out.
Reference: amazon.com/Thinking-Systems-…
Alexios @Mantzarlis, former Google Trust & Safety has since been in the trenches with Indicator finding scams and deception online. A few quotes from our conversation (link below):
* Fix products, don't just moderate: "I'll die on the hill that at most platforms there are still people trying to fix things and do things right. The question is: (a) can you get to them, and (b) do they have the agency to intervene? Taking things down or blocking content is always easier than proactively changing products and going through all the launch processes."
* Meta: "Meta obviously has a more interesting and in some ways worse track record on some things, but it's far more reliable at actually getting back to us on what we found, and often taking things down."
* Journalism as prioritization: "Covering platform deception pushes it further up the triage list for decision-makers inside the platform, because they have these infinite inboxes and they're looking for shortcuts on which decision to make. Making noise helps things get fixed."
* More people needed: "The platforms have the resources, the means, and the sophistication to do a lot of this already. Part of it is a triaging question — there's just so much. But I think the answer is: hire more people. Instead, we're in the midst of an ideological pushback against having humans. And it doesn't mean they can't be AI-empowered humans — it's not either/or."
Here is the show page with transcript (AI crawlable) and the links to Spotify, Apple and YouTube video - wontfixpod.com/podcast/episo…
“The future not just of journalism but of responsible A.I., too, depends on preserving incentives for humans to produce the creative works on which a healthy society depends,” The Times wrote in its filing.
summary judgement briefs were due today in NYT/publishers vs OpenAI and MSFT
expect the judge to rule in days/weeks ahead whether this goes to trial.
(also, was surprised how flowery some of the legal writing was throughout the filings…)
nytimes.com/2026/09/04/techn…
Always good advice, but not new advice. threads.com/share/BAZG2X4i5e…
Rob Leathern retweeted
This problem hit someone in my family this week. Caused a lot of stress. So frustrating. Easy for tech experts, but scary to normal people.
Please tell your (older) family members this kind of thing is bullshit. It is a scam.
It has sound. It makes a desktop computer spin up. They should close or quit their browser and NOT call the toll-free number. It’s not Apple (or Windows) Support.
“Before explicit status codes like WONTFIX, engineers would either leave low-priority bugs open indefinitely or mark them as "Resolved" without explanation.
Creating WONTFIX provided a necessary third state: acknowledging that an issue is indeed a real bug, but making a deliberate product decision not to allocate resources to fix it (usually because it was too expensive, too complex, or an intended trade-off).”
wontfixpod.com
Got to give a lot of credit to @404mediaco for their early coverage of the Flock camera story
“…deception is not only widespread, it’s become industrialized” - @CraigSilverman
Fake AI person pitching an agency that allegedly works around tricky ads policies. Sigh.
