Anthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California's Broken Elections

The episode opens on a product launch that the hosts immediately convert into a governance argument. Fable 5, released on a Tuesday, tops nearly every benchmark, but its tokens cost twice as much as Opus 4.8. Jason Calacanis frames the cost question as unresolved: it is "unclear if this is going to alleviate some of the hand wringing around the cost of tokens," because the efficiency claim depends on the model needing fewer tokens in practice. The pricing detail matters less than what surrounds it. Calacanis notes that in April Anthropic declined to release Mythos publicly on the grounds that it had strong hacking capabilities, distributing it instead to a limited set of partners, and that Nikesh Arora of Palo Alto Networks vouched that "Mythos was the real deal" and had been used to seal vulnerabilities inside his own company.

The substantive complaint is not about benchmarks. It is about three practices disclosed around Fable 5: prompts are retained for at least thirty days, the model detects when a user is doing frontier AI research, and users in that category were silently downgraded. Calacanis stresses the concealment rather than the policy itself, noting the rule was "buried deep in a 319 page document" and that the company only partially retreated afterward, in Wired, with the statement that "We're changing Fable five safeguards for frontier LLM development to make them more visible." That wording is itself the tell: the change is to visibility, not to the practice.

Chamath Palihapitiya's answer sets the analytical frame the rest of the segment works within. He grants the model is "really incredible" and credits the lab for pushing the closed frontier, then separates two risks. The first is censorship: Anthropic has "essentially shown their hand, which is that they will increasingly take in prompts, evaluate the prompts, and decide what to do with them before they generate output to you." The second is enterprise governance risk, and he treats it as close to disqualifying, because the trigger is invisible to the buyer. A downstream scientist using the API, a business executive, a molecular researcher — any of them could trip a classifier and cut the company off from a capability it had built workflows around.

His conclusion is a procurement argument rather than a regulatory one: companies must start underwriting this phase of AI by asking who controls access, who is allowed to learn from the data, and whether they are accepting a single point of failure. His summary line about the lab is the sharpest thing said in the segment: "One thing I'll say about anthropic is they tell the truth." He then supplies his own caveat — "It's just that the truth sucks when you actually take it and you eat it" — which is a useful distinction. The criticism is not that Anthropic lies about its policy. It is that the policy, stated honestly, is one enterprises should refuse.

David Sacks escalates from disclosure to intent, and he brings a timestamp. Eight months earlier he had called Anthropic's approach "a very sophisticated regulatory capture campaign based on fear mongering," a take he says was considered spicy at the time and is now close to consensus. What he thinks Calacanis's summary missed is the scale of the breach of trust in the developer community, and he locates the breach precisely: "Remember, this is a company that said that it was against government surveillance. They are now retaining for thirty days every prompt and every output you send to one of these mythos class models. There are no exceptions."

The exceptions point is the one with commercial teeth. Enterprise customers who had signed zero data retention agreements do not get to keep them; their only option is to forgo the new model class. Sacks also widens the data at issue beyond prompts and outputs to the full context window — the memories, files, and accumulated material that agent platforms pass to models to improve responses. On his reading, retention exists to "build a profile on you, to classify you, and then to determine what capabilities it then unlocks." That is a claim about purpose, not just about storage, and it is the claim that turns a privacy complaint into a structural one.

The second grievance is product degradation. In Sacks's account, Anthropic reserved the right, "in anthropic sole discretion," to route a user to a lesser model in machine learning, AI research, and chip design research without saying so, to rewrite prompts in the background, and to charge full price for capability the customer did not receive. He calls this "completely anticompetitive" and describes the result as "a new level of AI haves and have nots." The walk-back, in his telling, covers only disclosure: users are now told when they are downgraded, but the downgrading continues.

Two examples carry the empirical weight. A user asking about mitochondria was downgraded; Ben Thompson of Stratechery asked "a very straightforward question about the relationship between cancer risk and GLP-1s" and was "kicked out." Sacks's inference is that the classifiers have an expansive view of who qualifies for demotion, and that the company is comfortable with that breadth. His verdict — "That is very Orwellian" — is rhetorical, but the underlying structural point is not: a private firm is adjudicating informational access by user identity, with no published criteria and no adversarial check.

Chamath adds the version of this argument that is harder to dismiss, because it does not require bad intent. Even a well-run lab will face incentives to favor counterparties. If a model provider has a strategic deal with Novartis and not Lilly, or with JPMorgan and not Citibank, there is an impetus to shape how people get information about competing products. He also raises the auditability problem: if a user believes an answer was manipulated, the provider should be able to produce "the trace of that model run-in that exact moment that generated that output," and he doubts the fingerprint will be preserved. That is a concrete, testable demand — model-run provenance — and it is the most constructive proposal in the segment.

Jason Calacanis offers the free-market counter, and it is a real one: early search engines sold paid inclusion, consumers noticed the results were worse, and they moved to the better provider. His question is why the same correction would not happen here. Sacks's answer is that the analogy assumes exit is available. "It's not good enough for them just to win in the free market. They're calling on the government to regulate and stop potential competitors and limit the number of models that you have access" — and, he adds, the regulation they want would apply to open source, which cannot practically be regulated, making it "a preemptive strike" against running models locally.

The most valuable part of the segment is Friedberg's first-hand account, because it converts an abstract argument about chokepoints into a specific workflow that stopped working. Ohalo does proprietary genomics work in agriculture: estimating the effect of genes and gene variants on living organisms, designing RNA guide sequences for gene-editing tools, and predicting whether a gene carries a phenotype. Six months earlier, the models could design a genetic construct to produce a specific protein "very simply and cleanly." Over the preceding weeks, that capability was restricted, on the theory that these tools create a bioweapon-type risk.