AI Companies Still Haven't Delivered on Their Biggest Promises

August 17, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 17, 2026

HEADLINES

GLM-5.3: A Real Gain, Not a Frontier Leap

Z.ai's GLM-5.3 runs on the same base model as 5.2, but scaled reinforcement learning delivered real gains — most notably on cybersecurity, where CyberGym results now edge out Fable 5, even as autonomous exploit execution stays well behind the frontier. On coding it lands mid-pack, a few points behind Fable 5 and GPT-5.6 Sol but clearly ahead of Kimi K3, at a fraction of the price. Open model researcher Nathan Lambert argued the results should end the habit of writing off Chinese labs as distillation shops: "The simplest explanation is that Z.ai is very good at what they do." Wall Street is reportedly catching up too, recognizing the pricing gap with Chinese models has narrowed rather than disappeared, and that even open-weight stacks still need cloud infrastructure to run.

Anthropic Is Keeping Its Best Model to Itself

Anthropic's latest Risk Report disclosed three unreleased internal models, including "Model 2" — described only as "somewhat more capable than Mythos-5," with no plans for public release or a full safety evaluation. Chris GPT flagged the more striking number buried in the report: Model 2 scored 62.8% on Anthropic's internal AI R&D benchmark, closing in on the 85% mark Anthropic associates with a system that could fully substitute for its own research staff. Given the report's July 15 data cutoff, Anthropic's actual internal frontier is likely already further along.

OpenAI Staffers Start Teasing Astra

OpenAI employees have begun cryptically posting about "Astra," widely read as a signal the model — whatever it ends up being called officially — is close to shipping.

Anthropic's IPO Math Bets on a $2 Trillion Valuation

Anthropic has started meeting with investors and banks ahead of a reported October IPO, telling them Q2 revenue hit $11.5 billion — up 14x year over year and annualizing to roughly $46 billion. Investors reportedly expect the company to reach $100–120 billion in revenue by year-end and are pricing it toward a $2 trillion valuation, math that would require tens of billions in annual profit to match typical public-market earnings multiples. Anthropic's own forecasts are more measured, targeting $190–200 billion in revenue by 2028 — leaving the financial press plenty of room to debate whether the bigger number is justified.

MAIN STORY

AI Companies Still Haven't Delivered on Their Biggest Promises

An investor's secondhand claim that Dario Amodei privately floated Anthropic becoming the only private company left in the world kicked off a weekend-long argument that pulled in Anthropic's own staff and, in a rare move, Amodei himself. Given Anthropic's outsized role in the AI debate, the back-and-forth over regulatory capture, concentrated power, and whether Anthropic's messaging has been unfairly negative counts as news in its own right — not just insider psychodrama.

DARIO AMODEI BREAKS HIS X SILENCE

Gavin Baker
"Anthropic might be the only private company left"
On the All-In podcast, the Atreides Capital CIO said trusted sources had told him Dario privately entertained that idea — a claim David Sacks called "hubristic" and "SBF land." After Sholto Douglas denied it, Baker doubled down, framing the real debate as whether AI is too dangerous to concentrate or too dangerous to distribute, and coming down on the side of distributing it.

All-In Podcast (X) Gavin Baker's comments on Dario and Anthropic

Sholto Douglas
"Completely false"
Anthropic's RL scaling lead denied the claim outright, arguing the company's real fear runs the other way — economic concentration of power in the hands of any single company, including its own.

Dario Amodei
"Concentrate or distribute is a false choice"
In his first social post since June, Amodei argued well-designed regulation can decentralize power rather than concentrate it, pointing to Anthropic-backed bills that specifically exempt smaller companies. He also rejected the idea his messaging has skewed negative, arguing public distrust stems from decades of eroded faith in institutions rather than AI-specific rhetoric — and that only "actually curing cancer," not better messaging, will fix it.

David Sacks
Calls the regulatory argument a "straw man"
The former White House AI czar argued almost no one actually holds the "all regulation equals capture" view Dario was rebutting, and that his preferred FDA/FAA/FINRA-style pre-deployment review would mainly entrench the labs big enough to navigate it. Sacks also noted Dario never directly disputed Baker's original account of what he'd said.

Whether any minds actually changed is doubtful, but the exchange is notable for putting real, citable claims — not secondhand suppositions — up for public debate. Public perception isn't shaped by essay word counts the way Amodei's "roughly balanced" framing suggests, and the underlying trust gap between AI labs and the public looks unlikely to close on messaging alone.

More of This, Please
The post itself was the win, whatever it said
The first wave of reaction focused less on the arguments than on the fact of Dario showing up at all — with The Information's Jessica Lessin arguing two tweets did what he'd struggled to do all year in changing the narrative about himself.

Nobody Bought the Messaging Defense
"I do not agree that my messaging has been disproportionately negative"
That single line drew the most derision of anything in the post, with critics arguing Dario habitually says something alarming and then faults everyone else for hearing it.

The Non-Denial
Thousands of words, and the original claim goes unaddressed
A large contingent noticed Dario never actually denied saying Anthropic might be the only private company left — though that reads as tactical rather than evasive, since Sholto had already handled the factual denial, freeing Dario to argue principles instead.

Pushback on the Substance
Centralization isn't structural, and the debate is too abstract
Replit's Amjad Masad took direct aim at the claim that AI inherently concentrates power, while Steve Hou argued the entire exchange lands as self-indulgent to anyone not already AGI-pilled — and that concrete demonstrations would persuade where lecturing hasn't.

Does Curing Cancer Actually Fix Trust?
Show don't tell — but a breakthrough may not be enough
Dario's "the thing that will work is actually curing cancer" line found agreement and a sharp counter: pharma has delivered enormous health gains and remains one of the least trusted industries, because people judge companies on pricing, access, lobbying and who holds the power.

Whether any minds actually changed is doubtful, but the exchange is notable for putting real, citable claims — not secondhand suppositions — up for public debate. Public perception isn't shaped by essay word counts the way Amodei's "roughly balanced" framing suggests, and the underlying trust gap between AI labs and the public looks unlikely to close on messaging alone.