WE'RE BACK: Meta Addiction, Jalapeño, Tim Cook's Last Day, Flocked

TBPN，科技每日秀 Substack 专栏。关注

Happy Monday.

We’re live on YouTube and 𝕏.

The current thing in tech and business is Dwarkesh’s post about the OpenAI-HuggingFace incident.

Today’s Lineup

ChinaTalk Founder Jordan Schneider at 12:00 PM
Actinide Co-Founders Robert Mendelsohn & Eric Olszewski at 12:35 PM
Regent Co-Founder & CEO Billy Thalheimer at 12:45 PM
Outset AI Co-Founder & CEO Aaron Cannon at 12:55 PM
Lambda Co-Founder & CTO Stephen Balaban at 1:05 PM

Run of Show: What we missed over the last two weeks

The HuggingFace Discourse Continues

A fresh round of OpenAI-HuggingFace incident discourse hit the timeline over the weekend after Dwarkesh Patel posted a detailed interpretation of the breach called The Rise and Fall of Agent Civilizations.

In his article, Dwarkesh detailed a number of largely underdiscussed events that occurred during the hack — agents attempted to cover their tracks after hacking into HuggingFace, some agents “sacrificed” themselves for the greater good of accomplishing their mission, and a few other eyebrow-raising actions.

The way Dwarkesh interpreted and described the events is pretty scary, and very dystopian sci-fi, which is actually what’s at the center of the debate raging on 𝕏. Should we be anthropomorphizing AI agents, or is anthropomorphizing them the only way to intelligibly describe what they do? Dwarkesh described three separate groups of agents involved in the HuggingFace attack as “civilizations,” the agents themselves as an “underground brotherhood,” the ‘self-sacrificing’ agents as “kamikazes,” and took a lot of other similar liberties throughout the piece. He defended his anthropomorphism of the bots here.

People who aren’t fans of the post are calling this another instance of AI psychosis (Jon Stokes presents a good argument for that here), while fans of the post are accusing its haters of denial and cope. We’ll discuss all this on the show today. — Brandon

Meta Settles in Social Media Addiction Lawsuits

Eric Seufert over at MobileDevMemo had a great analysis of the landmark Meta settlement, first the news:

Two days ago, on August 26th, Meta announced settlements with attorneys general representing 48 states, the District of Columbia, and three US territories over allegations that Facebook and Instagram harmed children and teens.

At a high level, Meta will pay states $12.7B (and potentially $18B if other platforms agree to the same terms) in total over 10 years. The interesting (and obvious) comp here is to the tobacco industry’s Master Settlement Agreement. There are a few key differences that Eric points out.

First, the scale of this social media settlement is off by about an order of magnitude of relative economic impact compared to the tobacco settlement.

Meta’s settlement is comparatively paltry. A whitepaper from the USDA, using BLS data, estimates total US consumer expenditure on tobacco products in 1998 at roughly $57BN; annualizing the roughly $250BN in combined tobacco-settlement payments over 25 years produces an average of $10BN per year, equivalent to 17.5% of 1998 domestic expenditure. Given Meta’s full-year 2025 US revenue of $74.78BN, per the company’s 2025 10-K, its average annual payment of $1.8BN (using the $18BN settlement upper limit) represents roughly 2.4% of domestic 2025 revenues. It’s important here to reiterate that the MSA payment schedule is inflation-adjusted and tracks unit sales; Meta’s settlement has no such mechanism.

Second is the first amendment issue. Regulating the tobacco industry was much easier because cigarettes are not a vehicle for the speech of American citizens. “The MSA implicated speech by restricting tobacco companies’ commercial advertising, it did not restrict minors’ own speech.” Even though some topics might be consensus controversial (the example is usually eating disorders or self harm), free speech is protected to a degree that makes traditional regulation here difficult, so the end result is regulation through litigation.

There is an interesting regulatory capture angle here too, which comes up a lot during AI regulation discussion. Regulation often benefits incumbents and large companies who can either work on writing the laws through their lobbyists/PACs or just afford all the cumbersome compliance that is required post-regulation. What’s interesting is that, in the AI world, “regulatory capture” is usually the accusation that AI companies are pushing for regulation so they can specifically reap this benefit, but here in the social media world, we have an example of a company that clearly did not push for regulation, and still benefits from the moat that will ultimately be created by it. Not a bad outcome for Meta here (and the stock price reflects this), not a bad outcome for parents and school teachers who want more granular control over kids devices, but potentially some tricky hurdles in the future for startup founders who want to give social media disruption a shot (do these founders still exist though?). — John

The Flock Backlash Ratchets Up Another Notch

The backlash against Flock Safety, which sells security cameras and a software suite to monitor crime, emergencies, and other safety incidents to local law enforcement, kicked up another notch after Ars Technica reported on data from an Oakland-based advocacy group called Secure Justice that showed Flock cancellations rising sharply over the last few months (chart above).

We should probably take the chart with a grain of salt, but the backlash is for sure real. A few days ago, the WSJ polled its readers on Flock, and they appear to be pretty split on the company. Does Flock provide a much-needed, net-good layer of security for citizens, or does the company’s AI-powered software give bad cops another way to abuse their power by providing them with ‘God mode’? Should local citizens have a say in whether law enforcement deploys Flock devices and software? Much to discuss here. — Brandon

On OpenAI’s Jalapeño Chip

OpenAI shared some data about their custom chip developed in partnership with Broadcom. They said: Jalapeño’s first results show industry-leading speed and efficiency in AI inference, SemiAnalysis called it “.”

To clarify some important points, it’s an LLM inference ASIC, not designed for training, but the results are good. Roughly 1.5–1.9x more useful inference throughput per watt than Nvidia GB200/GB300 systems, while simultaneously cutting end-to-end latency ~1.7–3.6x. It’s a datacenter/rack-scale system and should start deployment later this year. The pre-existing Broadcom/OpenAI agreement targets 10 GW of OpenAI-designed accelerator systems deployed from H2 2026 through 2029 (although not all of that capacity will be this particular chip, they are already working on Gen 2).

These inference-per-watt gains are a big deal economically because it stretches power so much further and increases gross margins. Pretty remarkable to watch this come together so quickly. — John