# AI Companies Still Haven’t Delivered on Their Biggest Promises
*The AI Daily Brief — Monday, 2026-08-17 · https://aidailybrief.ai/e/2026-08-17*

**AI's trust problem won't be fixed by messaging — only by delivering.**

A podcast rumor that Dario Amodei thinks Anthropic could be the last private company standing dragged the famously offline CEO into the public square. His response reframed the whole debate: regulation isn't automatically regulatory capture, AI structurally concentrates power regardless of rules, and the public's distrust is a decades-old crisis that no glitzy campaign can fix. The only thing that will work, he says, is actually curing cancer — and the fairest criticism of AI companies, his own included, is that they haven't yet delivered on their big promises to benefit the world.

---

## By the numbers
- **14x** — Anthropic's reported year-over-year revenue growth — $11.5B in Q2
- **$2T** — IPO valuation Anthropic investors told the FT they expect
- **$59-79B** — Annual profits a $2T Anthropic would need at average earnings multiples, per Fortune
- **$190-200B** — Anthropic's own reported revenue forecast for 2028
- **28.3%** — GLM 5.3 on Terminal Bench 3.0 — about five points behind the frontier
- **62.8%** — Anthropic's unreleased Model 2 on its internal AI R&D benchmark vs 50.3% for Mythos-5
- **1,000+** — Critical and high-risk vulnerabilities ZAI claims GLM 5.3 found in open-source repos
- **<1/10th** — GLM 5.3's per-token cost vs Fable 5 or GPT-5.6 Sol

## Headlines

### GLM 5.3 squeezes near-frontier performance out of a mid-sized model `[01:00]`
ZAI's new release is built on the same base as GLM 5.2, with the gains coming purely from scaling reinforcement learning. It lands about five points behind Fable 5 and GPT-5.6 Sol on Terminal Bench 3.0 (28.3%) but 11 points ahead of Kimi K3, and posts state-of-the-art results on agentic benchmarks like GDPVal — slightly inching out the US frontier.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-08-17#glm-53-drops

### ZAI built GLM 5.3 for cyber defense — and beat Fable 5 on Cyber Gym `[03:00]`
ZAI says cyber performance was a key focus of the RL run, noting that during the Hugging Face attack defenders were forced to use GLM 5.2 because frontier-model guardrails rendered them useless. In a WeChat post: "If the powerful attack ability is spreading, the defensive ability cannot be limited to a few closed source model companies." GLM 5.3 jumped seven points on Cyber Gym to overtake Fable 5.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-08-17#glm-cyber-defenders

### No, this isn't an open-source Mythos-level cyber weapon `[03:00]`
Despite the freakouts, being frontier-level at finding vulnerabilities is not the same as being frontier-level at exploiting them or autonomously executing attacks. ZAI is also taking a phased approach, testing with trusted partners before publishing the full weights.
*For: Eng, Legal*
Link: https://aidailybrief.ai/e/2026-08-17#not-a-cyber-superweapon

### Cheap on paper, mixed in practice `[04:00]`
On a per-token basis GLM 5.3 costs less than a tenth of Fable or 5.6 Sol, and early users are seeing roughly two-thirds the cost of Kimi K3 for the same task. But first real-world tests are mixed: ZAI claims 1,000+ vulnerabilities found in open-source repos and a privately reported Cursor vulnerability, while other testers found it behind Kimi K3 on game dev and painfully slow — possibly just release-day demand.
*For: Eng, Finance*
Link: https://aidailybrief.ai/e/2026-08-17#glm-cost-and-mixed-first-tests

### Nathan Lambert: stop being surprised by Chinese labs `[05:00]`
The open-model researcher argues ZAI's strength is post-training while Kimi is a pre-training masterpiece — and that the simplest explanation for GLM 5.3's results is that ZAI is very good at what they do. Writing off Chinese labs as mere distillers or benchmark-maxers, he warns, leads to underestimating what they're actually capable of.
Link: https://aidailybrief.ai/e/2026-08-17#lambert-stop-being-surprised

### Wall Street finally realizes Chinese models aren't frontier intelligence for pennies `[05:00]`
The pricing gap has contracted substantially: cheaper US models like Grok 4.6 and GPT-5.6 Luna are now cost-competitive with the best out of China. And per Morningstar's Malik Khan, even an enterprise consolidating on open-weight models still needs cloud infrastructure to run them — a tailwind for cloud providers, not the GPU-investment killer feared after the DeepSeek moment.
*For: Finance, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#wall-street-updates-priors

### Anthropic's best model is one you'll never use `[06:00]`
Anthropic's latest risk report disclosed three significant unreleased models, including "Model 2" — described as somewhat more capable than Mythos-5, with no plans for public release. On Anthropic's internal AI R&D benchmark it scores 62.8% versus 50.3% for Mythos-5, and the report's July 15 date means the internal frontier is likely already further ahead.
*For: Eng, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#anthropic-keeps-model-2-internal

### The gap between what we use and what the labs have is widening `[08:00]`
Between Anthropic holding back Model 2 and the new paradigm of government involvement in US frontier releases, the distance between publicly available models and internal lab capabilities is wider than it's been in the past — important context for any debate about how far behind Chinese models really are.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-08-17#internal-gap-widening

### OpenAI staffers are vague-posting about Astra `[08:00]`
Not wanting to be left out, OpenAI employees have started teasing Astra — suggesting the model, whatever it ends up officially being called, will be in our hands soon.
Link: https://aidailybrief.ai/e/2026-08-17#astra-vagueposting

### Anthropic's IPO pitch: 14x growth and a $2 trillion whisper number `[08:00]`
Anthropic has begun meeting investors and banks, telling them revenue hit $11.5 billion in Q2 — up 14x year over year, annualizing to $46 billion. Investors told the FT they expect a $2 trillion valuation (Anthropic itself reportedly hasn't discussed one), which would top SpaceX's debut and more than double the May raise. Investors see $100-120 billion in revenue by year-end; Anthropic's own reported forecast is $190-200 billion by 2028.
*For: Finance, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#anthropic-ipo-numbers

### The financial press starts poking holes in $2 trillion `[09:00]`
Fortune notes public stocks trade on earnings multiples, not revenue multiples — at an average multiple, a $2 trillion Anthropic would need annual profits of $59-79 billion. Then again, Anthropic isn't being valued as an average company, and with no clear precedent, all we have is opinion to be argued. Expect much more of this before the IPO.
*For: Finance*
Link: https://aidailybrief.ai/e/2026-08-17#two-trillion-debate

## Main episode

### The claim that started it all: Anthropic as the last private company `[14:00]`
On All In, Eutrades Capital CIO Gavin Baker said multiple people he trusts told him Dario Amodei has said Anthropic might at some point be the only private company in the world — just Anthropic, governments, and everyone else. David Sacks called the reported sentiment "getting into SBF land," and Baker said he'd discourage Dario from ever saying it again.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-08-17#baker-only-company-claim

### Completely false... In fact, one of the things we are most worried about is economic concentration of power. `[15:00]`
*— Sholto Douglas, Anthropic, responding to Gavin Baker's claim*
Anthropic's Sholto Douglas escalated the All In clip to X, calling the sourcing a lie fitted to a narrative and arguing the AI market is "literally the most competitive market in the world right now" — while conceding that with AGI, "capitalism gets super weird and what a company even is might look different."
Link: https://aidailybrief.ai/e/2026-08-17#sholto-completely-false

### Baker's real argument: too dangerous to concentrate, or too dangerous to distribute? `[16:00]`
Baker's follow-up said the rumor is believable because it's consistent with Dario's public messaging, and framed the core question as whether AI risk is best handled by concentration via regulation or wide distribution — siding with Zuckerberg that extreme concentration of power is "inherently problematic." He declared Dario had lost the regulatory argument and urged him to become a more positive advocate for his own industry.
*For: Exec, Legal*
Link: https://aidailybrief.ai/e/2026-08-17#concentrate-vs-distribute

### Dario breaks his silence: regulation ≠ regulatory capture `[19:00]`
In a rare post — he follows zero accounts and last posted in June — Dario called the concentrate-or-distribute framing a false choice, arguing the Silicon Valley shorthand of regulation-equals-capture underrates "the decentralizing power of objective and fair institutional processes." He noted Anthropic's proposals deliberately exempt smaller players, and said he's supportive of the Trump administration's reported pre-deployment testing approach and Demis Hassabis's FINRA-like entity idea.
*For: Legal, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#dario-regulation-not-capture

### Dario: AI concentrates power because of scaling laws, not regulation `[21:00]`
His structural argument: AI tends to concentrate power for reasons rooted in the extreme implications of scaling laws. Open weights help some, but merely shift the concentration to whoever has the most compute and chips — which is roughly the frontier labs plus hardware providers.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-08-17#dario-scaling-concentrates-power

### By far the most accurate criticism of AI companies, including Anthropic, is that we haven't yet delivered on our big promises to benefit the world. `[23:00]`
*— Dario Amodei, CEO of Anthropic, on X*
Dario's diagnosis of AI's trust problem: a decades-old crisis of trust in companies, governments, and tech that no glitzy marketing campaign can fix. "Saying that AI will cure cancer is more of a cliché than it is inspiring... The thing that will work is actually curing cancer." He says Anthropic is ramping up in biology and medicine, with "early glimmers in the coming months."
*For: Exec, Marketing*
Link: https://aidailybrief.ai/e/2026-08-17#dario-havent-delivered

### Did Dario change the narrative? Depends who you ask `[24:00]`
The Information's Jessica Lessin said two tweets did what Dario struggled to do all year — change the narrative and deliver an accessible message. But a loud chorus zeroed in on his claim that his messaging hasn't been disproportionately negative, accusing him of gaslighting and of never actually denying the original "only private company" claim.
*For: Marketing*
Link: https://aidailybrief.ai/e/2026-08-17#reactions-split

### Why Dario and his critics keep talking past each other `[26:00]`
PR expert Lulu Cheng Meservey's clinical read: when accused of negative messaging (a qualitative charge), Dario rebuts that it's balanced because he's written one essay on benefits and one on risks (a quantitative defense). He measures analytically; everyone else goes off vibes — so this won't be the last time. She also noted the tactical sequencing: Sholto did the fact-checking first, letting Dario come in at the level of principles.
*For: Marketing, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#lulu-quantitative-vs-vibes

### On messaging, Dario is just wrong `[26:00]`
You can't claim to understand that social media clips the most negative parts and then do an endless string of interviews full of easily sound-bitable negative statistics. Public perception isn't shaped by a word count of positive versus negative essays — arguing otherwise shows a radical misunderstanding of the media environment a leader of this significance has to operate in. The trust gap is real and messaging alone can't fix it, but you still have to understand how messaging plays into it.
*For: Marketing, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#nlw-dario-just-wrong-on-messaging

### Scaling laws are not laws of physics `[28:00]`
Replit's Amjad Masad pushed back on Dario's centralization argument: 125 years of super-exponential gains in compute price-performance, plus algorithmic and hardware efficiency improvements, mean there's no reason to assume AGI-level capabilities will always require a data center. Scaling laws are empirical relationships for particular architectures and datasets — change any factor and you get a different curve.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-08-17#masad-scaling-laws-not-physics

### Curing cancer won't fix the trust problem either `[29:00]`
OpenAI's Angel Brodin countered Dario's show-don't-tell thesis: pharma delivered some of the greatest improvements in human health and remains one of the least trusted industries. People will judge AI companies by pricing, access, lobbying, opacity, how gains are distributed, and who holds the power — and if distrust of powerful institutions is the root cause, the answer can't be asking people to trust a few powerful institutions even more.
*For: Exec, Product*
Link: https://aidailybrief.ai/e/2026-08-17#brodin-trust-beyond-breakthroughs

### A few hundred words on X might be Dario's best medium `[30:00]`
Nothing got resolved, but debating what was actually said beats debating suppositions — and these posts prove there's value in participating in social media even for someone who hates it. Dario's 13,000-word essays and long interviews get strip-mined for negative soundbites; a few hundred words on X is much harder to take out of context, and the public gets to participate in a conversation whose stakes belong to everyone.
*For: Marketing, Exec*
Link: https://aidailybrief.ai/e/2026-08-17#nlw-public-discourse-takeaways

*Today's sponsors: KPMG, Blitzy, Robots and Pencils, Hyperagent — offers at https://aidailybrief.ai/sponsors*

---
Transcript: https://aidailybrief.ai/e/2026-08-17/transcript.md
Listen: https://pod.link/1680633614 · Ad-free: https://patreon.com/aidailybrief
© 2026 The AI Daily Brief — Until next time, peace ✌