Issue #3

WHAT IS REAL. 20-second clips with audio generated in the same pass, from text or images, with keyframe control. Live via API since August 4. This part is genuinely impressive.

WHAT IS MARKETING. 'HD' means 720p native plus an upscaler, and nobody says what resolution survives at the full 20 seconds. Every benchmark win is a BFL-run preference test with no sample sizes. The two models everyone actually measures against, Veo and Sora, appear in none of them.

THE PROMISE. 'open weights dropping soon' is an undated roadmap line, for a distilled Dev variant, scheduled last in the rollout. To their credit, BFL shipped FLUX.1 weights before. To reality's credit, 'later in 2026' is not 'soon'.

THE NUMBERS, CHECKED. every stat matches the report. Heavy users up from 13 to 39 percent in three years. Refuseniks down to 4.3. And 43.5 percent admit AI is their shortcut, not their tutor. Self-reported, so treat that as a floor.

WHAT THE HEADLINE UNDERSOLD. this is 483 self-selected business majors at a single school, and the authors disclaim generalizability themselves. Meanwhile the UK's national survey has student AI use at 94 percent. Kogod is not the surge. Kogod is the lagging indicator.

THE STAT THAT ACTUALLY MATTERS. interview questions about AI skills nearly quadrupled in two years, to 42.6 percent. Measured by asking students, not employers, but the direction is unmistakable. The question moved from the classroom to the hiring loop.

THE CHARGE. DOJ says OpenAI's green-card recruitment was built not to find American applicants. PERM jobs missing from the public careers site. Applications accepted by paper mail only. Ads placed on radio, late at night. Subtle.

THE FINE PRINT. fewer than ten positions at issue. $1.2 million penalty plus a $2 million back-pay fund, plus three years of DOJ monitoring. OpenAI disagrees with the findings and paid anyway, which is how every one of these agreements ends.

THE PATTERN. this is the 13th settlement under the DOJ's revived initiative. Facebook paid $14.25 million in 2021. Apple paid $25 million in 2023. The recruitment playbook is an industry standard. So is the settlement.

WHAT WAS ACTUALLY SAID. one aside on a podcast. 'SSI says that they'll come out with their model in August.' Gavin Baker, an investor who appears on none of SSI's disclosed rounds, quoting an 'SSI says' that nobody can locate.

WHAT SSI SAYS. nothing. No date, no model, no event. The website still reads like a vow of silence: 'SSI is our mission, our name, and our entire product roadmap.' The company was founded on the explicit promise of no products before superintelligence.

WHY THE RUMOR LIVES ANYWAY. Nvidia reportedly wired about $5B into SSI last month, and Sutskever now allows that 'gradual release would be part of any plausible plan'. Money that size rarely stays patient and silent.

THE TEST. August has three and a half weeks left. Either a model shows up, or this joins the long list of launch dates that only ever existed on podcasts.

Follow the money in a circle.

Nvidia invests in Volta. Volta buys Nvidia chips. Michael Dell's family office invests in Volta. Dell supplies the hardware. The same dollar counts as demand every time it passes a familiar hand.

The $10 billion? A compute commitment spread over roughly six years, for capacity that does not exist yet: a 133-megawatt site in Norway, leased from a Bitcoin miner, not fully deployed until March 2027.

The customer? Volta's own press release names only 'an AI Lab'. The word Anthropic comes from Bloomberg's anonymous sources. Anthropic declined to comment.

Best detail in the whole story: Bitdeer, the Bitcoin miner renting them the barn, jumped 14 percent on the announcement. The market is pricing the story, not the steel.

WHAT THE REPORTING SAYS. Axios and CNN, via anonymous sources, report open-weight models are exempt from the new pre-release review. The executive order behind the framework never mentions open weights at all. The framework document itself is unpublished. 'Reportedly' is doing the heavy lifting here, and it is getting tired.

THE LOGIC, STEELMANNED. you cannot recall open weights after release. Reviewing closed models is where government access adds something new. And reviewing American open models while Chinese ones sail past would be unilateral paperwork.

THE LOGIC, INVERTED. open weights are precisely the models you cannot patch later. Pre-release is the only intervention point that will ever exist for them. The framework exempts the one case where review has teeth.

AND THE KICKER. participation is voluntary anyway. A voluntary review with an exemption is a press release with extra steps.