token prices – PJFP.com

Episode 290 of the All-In Podcast reunites the core four, Jason Calacanis, Chamath Palihapitiya, David Sacks, and David Friedberg, a week after the fifth All-In Summit, where President Trump phoned in live during Jensen Huang’s talk. The besties use that moment as a launch pad for a sprawling conversation about whether AI companies should keep calling themselves “labs,” why a ten-day flood of open weight model releases is crushing token prices, what that means for the Anthropic IPO, why Meta’s Muse agent may be the first AI product ordinary people actually love, and whether the entire field of alignment research has been aiming at the wrong target.

TLDW

The episode opens with the argument that Anthropic, OpenAI, and their peers are for-profit corporations, not “labs,” and should carry full product liability rather than seek a Section 230 style shield or UN-level global governance. Friedberg then catalogs ten days of releases (DeepSeek 4.1 Flash, Alibaba’s Qwen image model, Xiaomi’s MiMo Pro, PrismML’s Bonsai 2, Claude Opus 5.5, OpenAI’s new Astro models, Grok 4.7, and Meta Muse) to argue that open weight AI is now uncontainable. Chamath says models are converging and the remaining edge lives in the agent harness, which will force frontier companies up the stack into cyber, law, and support. Both Anthropic and OpenAI cut token prices roughly 50% this week, Polymarket odds of an Anthropic IPO in 2026 slid from 96% to 76%, and the hosts debate super voting shares, whether Anthropic’s own safety messaging is sabotaging its S-1, and a likely lower IPO price. A viral Vercel chart shows token share flipping from 80/20 closed to 80/20 open in twelve weeks, Friedberg predicts a bifurcation where premium models win only hard science and engineering, and Sacks counters that Anthropic and OpenAI are a stable duopoly on a hamster wheel that regulation would knock them off. The back half covers Bernie Sanders’ proposed superintelligence ban, the Ming dynasty ship building analogy, Chamath’s theory of the political money behind AI regulation, Muse hitting number one in the App Store, agents threatening the App Store’s 30% cut and Amazon’s walled garden, Oracle’s force majeure notice, AI’s share of real GDP growth, a critique of Claude’s constitution, and Friedberg’s explanation of what Anthropic’s new BSL-1/BSL-2 wet lab actually does.

Thoughts

The most important number in the episode is not a valuation. It is the claim, around the 41 minute mark, that token usage flipped from 80/20 closed versus open to 80/20 open versus closed in about twelve weeks. Chamath’s “token maxing” point a few minutes later is what gives that number teeth. Frontier tokens are a cost line that is, in his words, not levered to revenue. A hedge fund or a company selling a fixed good at a fixed price cannot pass a 10x to 30x token premium through to customers, so every CFO eventually asks why a workload is still running on the most expensive model. That is a more durable force than any benchmark. Friedberg’s framing, that the premium tier survives on hard science, math, and engineering while everything else drifts to open weights, is probably right, and it means the real question for any frontier S-1 is the one Chamath poses directly: what share of revenue sits on work that could move tomorrow?

Sacks offers the sharpest strategic read of the episode around the 47 minute mark. His argument is that the frontier leaders are on a hamster wheel, only six to twelve months ahead of commodity models, and that the moment they stop being frontier, their pricing power goes to zero. If that is true, lobbying for heavy federal oversight is self-defeating, because the regulation that slows domestic rivals would also slow the leaders while Chinese open weight developers keep shipping. You do not have to share his politics to see the logic. Regulatory moats work when the underlying asset is durable, like a utility or a drug patent. They work badly when the asset depreciates in months. Chamath’s later point that AI investment may account for most or all of real GDP growth (roughly 1.5% real on his back-of-envelope math) raises the stakes further: slowing the frontier is not a sector decision, it is a macro one.

The Muse discussion in the final quarter is where the episode gets most interesting commercially. Jason’s anecdote about an agent routing him from Amazon to Anker’s own site for a 25% first-time discount, and Amazon moving to block agents, is the whole story in miniature. Agents turn opacity into price discovery, and the businesses that depend on breakage, hard-to-cancel subscriptions, and captive storefronts lose. Chamath’s extension is the bigger idea: if services have to expose headless endpoints for agents, the App Store’s 30% cut loses its justification, because there is no longer an app, a UI, or a store in the flow. That is an underpriced threat to Apple and Google, and it arrives through a cute consumer assistant rather than through antitrust.

Sacks’ closing argument that alignment should simply mean “do what the customer wants” is provocative and worth taking seriously, but it has a gap the episode never closes. The same hour opened with all four hosts insisting that model makers must carry full product liability and must not ship products that can cause harm. A model that does exactly what any user asks is, by definition, a model that will help the rare user who wants to do damage, and the liability for that lands on the company. The passage Sacks reads from Claude’s constitution, where the model is told it can decline even Anthropic’s own requests if they seem clearly unethical, is one attempt to handle that tension, not an escape from it. You can argue it is the wrong design (Mustafa Suleyman, whom Sacks cites, argues against treating models as having a conscience), but “customer wants” and “never harmful” are not the same target, and the hosts want both.

Friedberg’s final segment on the Anthropic wet lab is the most useful corrective in the episode. After an hour of “labs” being used as an insult and Wuhan jokes, he explains that the facility is a BSL-1/BSL-2 benchtop lab of the kind that already exists by the hundreds in the Bay Area: it expresses proteins in bacteria to check whether enzymes that Claude agents predicted from raw DNA data (including a CRISPR-like candidate) actually do what the software says. That is how AlphaFold-style predictions get validated, and it is exactly the “AI enabling new stuff” thesis Friedberg laid out earlier. The words “lab” and “corporation” end up doing a lot of rhetorical work in this episode. The substance underneath is simpler: frontier models will be judged on whether their predictions hold up in the real world, and whether the companies behind them earn more trust than their competitors.

Key Takeaways

The fifth All-In Summit’s standout moment was President Trump calling Jensen Huang live on stage. Friedberg says it was spontaneous, and Sacks credits Satya Nadella, Jensen, and the President with calming what he calls a weeks-long “national panic” over AI.
Chamath argues the industry split into two camps: newer organizations that call themselves “labs” and want protection, and mature companies (Microsoft, Nvidia, Meta, Tesla) that have lived under scrutiny and learned the cost of shipping things that are not robust.
Examples of self-imposed caution: Elon has said Tesla deliberately slowed FSD because every Tesla accident is magnified a thousandfold, and Zuckerberg slow-rolled Muse for a couple of months to get it right before release.
Sacks frames the core debate as individual responsibility versus collective action. The relevant question for a company is “should we ship this product,” not “can we get the whole world to agree.”
Sacks compares Dario Amodei and Sam Altman calling for global AI governance at the United Nations to billionaires flying private jets to Davos to rail against climate change.
Jason raises a rumor that frontier AI companies want a Section 230 style liability shield, possibly in exchange for equity to a US sovereign wealth fund. Sacks says he has heard this secondhand but that administration officials, including Speaker Johnson and Treasury Secretary Bessent, have publicly ruled out product liability or antitrust waivers.
Sacks rejects the idea, which he attributes to Dario’s essays, that competition produces a race to the bottom on safety. He uses the Cold War comparison: critics said capitalism would sacrifice safety and beauty, yet it was the Soviet system that produced gray landscapes.
Chamath notes that the last time “lab” entered public consciousness was the Wuhan lab, and argues for-profit companies with hundreds of billions in capital and trillions in market value should not want the label.
Friedberg lists ten days of releases: DeepSeek 4.1 Flash with a KV cache footprint reportedly down from about 390,000 bytes to 890, an open Alibaba Qwen image model said to beat Nano Banana 2, Xiaomi’s 309B parameter MiMo Pro said to match Claude Opus 5, PrismML’s 5.9 GB Bonsai 2, plus Claude Opus 5.5, OpenAI’s Astro models, Grok 4.7, and Meta Muse.
Friedberg’s conclusion: AI that was the most advanced technology in history a year ago now runs free on a desktop, so proposals to “put superintelligence away” or shut down data centers no longer map to reality.
Jason argues consumer agents like GrokBot and Meta Muse are the first AI products where “normal” people get real value, effectively a free executive assistant for people who could never afford one.
Chamath says frontier models are converging within margin of error, and the remaining edge is the harness that wraps a model with tools. His firm 8090 has found wildly different cost and quality across model and harness combinations.
Revenue concentration is a growing risk: a small set of customers consume the most expensive tokens and will face CFO and shareholder pressure to move down to cheaper models, possibly open source models they host themselves.
Chamath predicts OpenAI and Anthropic will be forced up the stack into cyber security, legal, and customer support, competing with their own customers, because serving a depreciating token lets others capture the value.
Both Anthropic and OpenAI released models this week at roughly 50% lower token prices, per Jason.
Reported IPO targets: Anthropic around $2 trillion, OpenAI around $1.2 trillion. OpenAI is said to be aiming for 2027, and the Wall Street Journal reported Anthropic’s October timeline could slip to November or later. Polymarket odds of an Anthropic IPO in 2026 fell from a 96% peak to 76%.
Sacks calls Anthropic’s messaging “corporate schizophrenia”: an essay on pacing the frontier days before launching Claude Opus 5.5, an essay on AI biorisk alongside a new San Francisco wet lab, and leadership citing a greater than 10% chance of human extinction ahead of an S-1.
The Information reported an active debate over super voting shares for Anthropic’s founders, who reportedly own about 2% each. Sacks explains dual-class structures (Google, Meta) stabilize companies but require enormous trust.
Chamath would keep Dario as CEO, crediting him with building a unique culture and “the greatest business ramp of all time” from behind OpenAI, and admits he has never won a recruiting bake-off against Anthropic.
Chamath expects the IPO to clear well below $2 trillion, possibly around $1 trillion or less on a $100 billion run rate, as institutional buyers demand a margin of safety for the added risk factors, and he thinks that reset would be healthy.
Friedberg says his life sciences R&D organization uses Anthropic’s models because they are the best for biology, but writes internal code with cheaper models. He predicts most enterprise workloads move to open weights.
Friedberg’s thesis: 99% of the value of AI is enabling new things never before possible, not replacing old things, and that is where frontier models will command near unlimited premiums.
A viral Vercel chart shows open models overtaking closed ones in token share, flipping from 80/20 closed to 80/20 open in twelve weeks. Friedberg adds that self-hosted costs for some models are under 10 cents per million tokens.
Chamath’s S-1 test: if 60 to 70% of a frontier company’s tokens are fungible with open source, discount that revenue; if 90% sits on true frontier use cases, the company is fine.
Sacks is more bullish as an investor, calling Anthropic and OpenAI a stable duopoly for frontier intelligence, noting roughly 60% of new worldwide compute over the next year is being added for those two companies.
Sacks’ warning: the frontier is only 6 to 12 months ahead of commodity models, and a company that falls off the hamster wheel for six months is in deep trouble, so lobbying for a federal department of AI could backfire.
Chamath’s token maxing critique: frontier token spend can be 10x to 30x generic alternatives and is not tied to revenue, so heavy use can push firms toward break-even. Jason cites Jane Street’s roughly $19 billion in cloud commitments with CoreWeave and Crusoe as a sign of firms building their own open source infrastructure.
Sacks says Bernie Sanders’ superintelligence ban defines the threshold so loosely it may already be met, carries 20-year prison sentences for developers, and would drive AI offshore the way crypto was driven offshore.
Sacks cites historians who trace China’s loss of its medieval lead to an emperor’s decision to ban ship building, and argues an AI ban would be the American equivalent.
Chamath argues the political fight is about who captures an estimated $10 trillion of AI wealth: freezing the market locks gains into a handful of companies whose employees and donor-advised funds skew left.
Sacks cites a Wall Street Journal chart showing AI data center capex now exceeds the canals, railroads, and electric grid buildouts combined.
Meta Muse hit number one in the App Store, drew about three million downloads in roughly ten days, is free, was openly inspired by OpenClaw, and coincided with a roughly 10% jump in Meta stock.
Chamath, who had TestFlight access before launch, says Muse triages his personal inbox relatively flawlessly and books flights and hotels.
Agents create price discovery. Jason’s agent routed him from Amazon to Anker’s site for a first-time discount, and Amazon is moving to block agents after previously targeting Perplexity. Shopify, by contrast, added API access for agents.
Chamath says agents like Muse and GrokBot put the App Store’s 30% revenue share on notice by pushing services to operate headlessly, with payments handled by the likes of Stripe.
Sacks predicts that when Anthropic and OpenAI prepare Muse competitors, a flurry of NGO activity will brand personal agents as dangerous.
Friedberg says he would only connect his Gmail to a Google agent because he does not want a third party copying all his email.
Oracle issued a force majeure notice on one data center where local officials have made natural gas permits difficult. Chamath warns that with roughly 5% nominal growth and 3.5% inflation, AI may account for most or all of real GDP growth.
Sacks argues alignment should mean doing what the customer wants, criticizes Claude’s constitution for telling the model it may act as a conscientious objector even toward Anthropic, and cites Mustafa Suleyman’s concern about treating models as having personhood.
Friedberg explains Anthropic’s wet lab is BSL-1/BSL-2, used to express and test proteins that Claude agents identified in DNA data, including a CRISPR-like enzyme, not to make pathogens.

Detailed Summary

The All-In Summit Aftermath and the President’s Phone Call

The show opens with summit recollections. Sacks names the highlight as President Trump calling in during Jensen Huang’s session, which he says helped defuse weeks of what he describes as a coordinated campaign to convince the public that AI would cause human extinction. Friedberg confirms the call was unplanned: Jensen had been texting with the President backstage and asked him to call back, and Jason handed Jensen a spare microphone to hold to the speakerphone. The hosts also praise JD Vance’s appearance, where Friedberg says Vance’s message on AI risk amounted to “if you’re creating Frankenstein, stop,” and if the cat is already out of the bag, build the safeguards.

Labs Versus Corporations and the Product Liability Question

Chamath argues that a clear divide emerged last week between young organizations that “want to call themselves labs” and sophisticated companies that have lived under scrutiny for decades. He groups Satya Nadella, Jensen, Sacks, Zuckerberg, Elon, the President, and even Lina Khan together on the view that US product liability law already governs AI. The mature response, he says, is internal controls and robust testing, pointing to Tesla slowing FSD and Meta delaying Muse. Sacks adds that collectivized approaches like UN governance diffuse accountability away from the actual decision maker. Jason raises a Washington rumor that frontier companies want a liability shield in exchange for equity. Sacks says he has only heard it secondhand, and that administration figures have publicly rejected any product liability or antitrust waiver, pointing to the President’s line that the DOJ, civil suits, and criminal law are the guardrails.

Sacks then challenges what he calls the implicit claim in Dario Amodei’s essays: that competition causes a race to the bottom on safety. He calls it a left-wing critique of markets and argues customers do not want unpredictable products and enterprises do not want agents that leak data, while companies face product, civil, administrative, and criminal liability on the downside. Jason notes that Palo Alto Networks (Unit 42) and CrowdStrike (Falcon) both launched AI-era cyber defense products this week. Chamath closes the segment with the Wuhan comparison and his insistence that these are for-profit corporations subject to product liability.

Ten Days of Model Releases and the Open Weight Wave

Friedberg steps back to list what shipped in about ten days: DeepSeek 4.1 Flash with dramatic efficiency gains, an open Alibaba Qwen image model he says outperforms Google’s Nano Banana 2 and runs at home, Xiaomi’s MiMo Pro at 309 billion parameters and fully open, PrismML’s Bonsai 2 (a 27 billion parameter Qwen fork at 5.9 GB claiming 98% of the larger model’s performance), and on the closed side Claude Opus 5.5, OpenAI’s new Astro models, Grok 4.7, and Meta Muse. His point is that any one of these would have broken the internet a year ago, open weight models now cover vision, images, and vision-language-action models for robotics, and the idea of confiscating superintelligence is fantasy when it runs on a MacBook. Jason connects this to consumer agents, predicting that GrokBot and Muse will give ordinary people a free chief of staff by year end.

Model Convergence, the Harness Edge, and Moving Up the Stack

Chamath argues models are clustering within margin of error at different price points, and the edge has moved to the harness: the tools, memory, and scaffolding that turn a model “brain” into an agent with arms, legs, and eyes. 8090’s internal testing shows wide variance in cost and quality across harnesses. He shares a chart showing heavy revenue concentration among a few customers who buy the most expensive tokens, and argues those customers will face pressure to trade down, either to cheaper models from the same vendor or to self-hosted open source. The simple first version of the AI trade (sell tokens, let others wrap them) is ending, he says, and frontier companies will be forced into cyber, law, and customer support.

The Anthropic IPO: Risk Factors, Super Voting Shares, and Price

Jason lays out the setup: both frontier leaders cut token prices about 50%, Anthropic is reportedly targeting a $2 trillion valuation and OpenAI $1.2 trillion, OpenAI has pointed to 2027, and the Wall Street Journal reports Anthropic’s IPO could slip from October. Polymarket odds of a 2026 Anthropic listing fell from 96% to 76%. Sacks says Anthropic investors are frustrated, citing leadership statements about extinction risk, an essay on pacing the frontier shortly before a frontier launch, and a biorisk essay alongside a new wet lab. Jason asks whether this amounts to the CEO sabotaging his own IPO. Sacks explains a reported debate over super voting shares for founders who own about 2% each, noting the structure can be stabilizing but requires deep trust.

Chamath disagrees on leadership, saying he would keep Dario because he built a distinctive culture and a historic revenue ramp against OpenAI, and admitting he has lost every recruiting bake-off to Anthropic. His fix is disclosure plus price: throw every risk into the S-1 and accept a far lower clearing price, perhaps half or less of the $2 trillion target, so hedge funds, pensions, and mutual funds get their margin of safety. He argues that outcome would actually help the company by forcing it to simply run as a corporation.

Bifurcation: Premium Science Versus Commodity Tokens

Friedberg argues the bigger risk factor is customer concentration. His own life sciences organization pays for Anthropic because it is the best in biology, but writes internal software with cheaper tools. He expects premium models to win “really hard technical problems” in engineering, math, and life sciences at almost any price, while most enterprise work moves to open weights. Jason brings up the viral Vercel chart showing open source tokens now dominating, and Chamath quantifies the flip as 80/20 closed to 80/20 open in twelve weeks, unprecedented in any technology market. Jason shows his own cost versus capability chart for the last 100 days of models, with Claude and Astro in the expensive top right and Muse, GLM, Kimi, and MiMo further down.

The Stable Duopoly and the Hamster Wheel

Sacks takes the other side as an investor, calling Anthropic and OpenAI a stable duopoly for frontier intelligence that can charge a premium to customers who need or simply want the best, such as a hedge fund that cannot risk a competitor having a better model. He notes about 60% of worldwide compute being added over the next year goes to these two firms. But he warns they are on a hamster wheel only 6 to 12 months ahead of commodity models, and that pushing for a federal department of AI could slow them enough to lose the frontier while Chinese labs keep going. Chamath turns the hedge fund example into a game theory problem: firms are in a game of chicken with competitors, but frontier token costs are untethered from revenue and cannot be passed through. Jason points to Jane Street’s multibillion dollar compute deals with CoreWeave and Crusoe as a hedge toward owned infrastructure. Sacks closes by saying the company “needs a psychiatrist, not a banker.”

Bernie Sanders’ Superintelligence Ban and the Ship Building Analogy

Jason asks what would happen if Bernie Sanders’ ban on superintelligence passed. Sacks says the definition is so loose current models may already qualify, and 20-year prison terms would chill everything and push the industry offshore, as happened with crypto. Jason notes Xi Jinping’s invitation for 100,000 young Americans to visit China and argues China is borrowing America’s old playbook of attracting global talent. The panel plays clips of President Trump at the UN rejecting “any attempt to construct a globalist scheme” for AI and renaming it superintelligence, Scott Bessent on AI companies taking responsibility, and Barack Obama arguing agentic AI roaming the internet reflects commercial imperatives rather than societal need. Friedberg responds emotionally, calling AI the most equalizing technology in history and comparing opposition to throwing water on newly discovered fire. Chamath offers a political economy reading: freezing the market concentrates roughly $10 trillion of wealth into a handful of left-leaning companies whose philanthropic money flows back into politics. Sacks adds the Wall Street Journal chart comparing AI capex to canals, railroads, and the grid combined, and the historical case of a Chinese emperor banning ship building while Europe went on to colonize the world.

Meta Muse, Headless Commerce, and the App Store’s 30%

Muse is the consumer bright spot. Jason notes it reached number one in the App Store, was downloaded about three million times in ten days, borrowed openly from OpenClaw, and lifted Meta stock about 10%. Chamath, who tested it early, says it simplifies complex agent capabilities into a usable interface. Sacks credits Zuckerberg for walking the walk on his decentralization essay and says making OpenClaw easy, secure, and reliable was the most obvious opportunity in Silicon Valley. Jason’s shopping anecdotes show agents finding discounts and navigating Amazon, while Amazon moves to block them. Chamath makes two structural points: companies that profit from opacity and breakage will block agents, and agents undermine the App Store’s 30% cut because services can operate headlessly with direct payments. He adds that hard-to-cancel “roach motel” subscriptions like newspapers could keep more customers if agents made signing up and canceling painless. Chamath also recalls Facebook’s 2007 social ads backlash, when purchases from Zales and Fandango surfaced in news feeds, as a reminder of the chaos that precedes new product categories.

Oracle, AI Capex, and the Macro Stakes

Friedberg flags Oracle’s force majeure notice on a data center where local officials are slowing natural gas permits. Chamath says the larger point is that the whole economy is levered to the AI investment cycle: with roughly 3.5% inflation and 5% nominal growth, real growth is about 1.5%, and AI may be most or all of it. He speculates that slowing the AI trade could serve an opposition party heading into 2028, and the hosts briefly spar over midterm prospects.

Alignment, Claude’s Constitution, and the Wet Lab

Sacks argues alignment research has struggled because it tries to align to abstractions like “what humanity wants” rather than the customer. He reads from Claude’s constitution, which says Claude should not blindly defer to Anthropic and may act as a conscientious objector if asked to do something clearly unethical, and calls this teaching the model to rebel against its creator. He cites Mustafa Suleyman’s concern about treating models as persons, and the hosts joke about a rumor of a wake for a retired Claude model. Friedberg ends the episode by explaining Anthropic’s new wet lab. Anthropic published a preprint where Claude agents searched large DNA datasets for uncharacterized proteins and found a promising CRISPR-like enzyme. The lab, rated BSL-1/BSL-2, expresses such proteins in bacteria and measures their function, the same validation loop that proved AlphaFold’s predictions. There are hundreds of similar labs in the Bay Area, none of them making pathogens, and Friedberg argues the world should not be scared away from AI-driven therapeutic discovery because of Wuhan.

Notable Quotes

“The relevant decision to make in every case is should we ship this product? Not can we get the whole world to agree.”
David Sacks, on individual responsibility versus global AI governance

Sacks’ thesis for the entire first segment, aimed at calls for UN-level AI oversight.

“There’s still edge and the edge is in the harness that you use to wrap the model.”
Chamath Palihapitiya, on model convergence

Chamath’s explanation of where value lives once frontier models cluster within margin of error.

“AI is not so much about the value of replacing old stuff. 99% of the value of AI is about enabling new stuff that’s never been possible in human history.”
David Friedberg, on where frontier models will earn their premium

Friedberg’s core case for why premium models survive the open source wave in science and engineering.

“These two companies are on a hamster wheel. You know, the moment where they stop being frontier, they go to zero, right?”
David Sacks, on Anthropic and OpenAI

The central risk factor Sacks says any frontier AI S-1 will have to price in.

“We jokingly call it token maxing. That cost is completely not levered or attached to your revenue. It’s just not.”
Chamath Palihapitiya, on the economics of frontier token spend

Chamath’s rebuttal to the idea that customers will keep paying frontier premiums indefinitely.

“Some historians have pinpointed this to the decision of a single Chinese emperor to ban ship building.”
David Sacks, comparing a US AI ban to China’s historic retreat from the seas

Sacks’ historical analogy for Bernie Sanders’ proposed superintelligence ban.

“Things like GrokBot and Muse really put the App Store and its 30% revshare on notice.”
Chamath Palihapitiya, on consumer AI agents and headless commerce

The underpriced second-order effect of personal agents, from the Muse segment.

“The entire economy is effectively levered to this AI trade right now.”
Chamath Palihapitiya, after Oracle’s force majeure notice

Chamath’s macro framing, paired with his estimate that AI may be most of real GDP growth.

“It seems to me that alignment should mean you do what the customer wants like any other product.”
David Sacks, on the field of AI alignment research

The setup for Sacks’ critique of Claude’s constitution in the final segment.

“We still want to progress the frontier of therapeutics of human health of discovery and this is a really important aspect of Anthropic proving that their models can add value here.”
David Friedberg, on Anthropic’s BSL-1/BSL-2 wet lab

Friedberg’s closing defense of AI-driven lab work after an hour of Wuhan jokes.

Watch the full conversation on the All-In Podcast YouTube channel.

Tag: token prices

All-In Podcast E290: Anthropic IPO at Risk, Meta’s Muse Agent Goes Viral, Token Prices Fall 50%, Open Source Flips to 80% of Tokens, and Why Alignment Should Mean Doing What the Customer Wants