DOGE kills its first bill, Zuck vs OpenAI, Google's AI comeback with bestie Aaron Levie

Chamath’s on-air statement in the Dec 20, 2024 episode is explicitly predictive: he and Jason talk about the upcoming holidays and say, “Three or four besties will be hitting the slopes… Three or four besties will be skiing together… We’re all going to be there, bro.” in reference to a ski trip the following week. 【6†turn3search0】

In the subsequent episode “2025 Predictions with bestie Gavin Baker”, released Jan 3–4, 2025, Jason opens by saying, “We’re all wrapping up our ski trip here,” and cues “found footage from the Bestie skiing.” The clip and banter clearly describe multiple besties skiing together: Jason narrates his own run, then Chamath’s (“There comes Champe… you really advanced this last season”), and then explicitly ranks the on-slope skiers as “me, Chamath, Friedberg in that ranking.” He praises Friedberg for having been “out there a number of days” and Friedberg confirms, “I think I skied four days.” 【5†turn2search1】 The episode description from several podcast directories also notes that the besties “welcome Gavin Baker and recap ski week,” placing this group ski trip in the immediate post-holiday window referenced on Dec 20. 【2†turn2search3】【2†turn2search6】

Putting this together: (1) on Dec 20 they predict that during the upcoming holiday period, 3–4 besties will ski together at the same destination; (2) in the very next regular episode after the holidays they explicitly recap a shared ski week, with at least three core besties (Jason, Chamath, and Friedberg) skiing on the same trip. That is sufficient to conclude that the prediction—“three or four besties will be skiing together” over that holiday period—came true.

Evidence from the episode transcript shows Jason proposing the experiment and explicitly saying Aaron Levie would be invited back in two weeks to see what happened if Box put 5% (~$30M) of its treasury into Bitcoin.(podscripts.co)

Reviewing the All-In episode list and transcripts around that window:

• The next regular episode after Dec 20, 2024 is “2025 Predictions with bestie Gavin Baker,” released Jan 3–4, 2025, which features only Gavin Baker as a guest; Aaron Levie is not listed as a speaker and does not appear in the transcript.(allin.onpodcastai.com)
• Aggregated episode indexes (Shortform, Metacast, Libsyn) show no other All-In episodes between Dec 20, 2024 and early January 2025 that include Aaron Levie.(shortform.com)
• Aaron Levie’s next appearance on All-In is “Trump's First 100 Days, Tariffs Impact Trade, AI Agents, Amazon Backs Down,” released May 2–3, 2025—more than four months later—and the episode description focuses on Trump’s first 100 days, tariffs, Amazon, and AI agents, not on Box allocating treasury funds to Bitcoin or reviewing such an experiment.(allin.onpodcastai.com)
• Public information about Box, Inc. (its Wikipedia entry and filings) contains no indication that Box adopted a Bitcoin treasury strategy in late 2024 or early 2025, which you’d expect if they had actually executed a $30M Bitcoin purchase of the type discussed on the show.(en.wikipedia.org)

Because (1) no All-In episode with Aaron Levie aired within approximately two weeks of Dec 20, 2024, and (2) his eventual return months later did not center on a Box-Bitcoin treasury experiment, Jason’s prediction that Aaron would be invited back within two weeks to discuss the impact of putting 5% of Box’s treasury into Bitcoin did not come true.

Evidence after the Dec 20, 2024 episode shows that both parts of Jason’s prediction materialized.

1. David Sacks returned as a host. In episode E208 on Dec 20, 2024, Sacks is absent and Aaron Levie is described as “the new fifth bestie,” with a chapter segment explicitly titled “Thoughts on Sacks’s new role,” indicating he had stepped away from regular hosting at that point. (allin.onpodcastai.com) On Jan 25, 2025, episode E212 is released with the explicit title “Trump’s First Week… + Sacks is Back!” and the speaker list shows David Sacks again labeled as a host alongside Jason, Chamath, and Friedberg. (allin.onpodcastai.com) Subsequent 2025 episodes (e.g., E218 on March 8, 2025 and later shows) continue to list Sacks as a host, confirming he returned and remained part of the core lineup. (allin.onpodcastai.com)

2. The show adopted an ongoing rotating ‘fifth seat’ for friends of the pod and newsmakers. Starting with the Dec 20, 2024 episode itself, the description explicitly frames Aaron Levie as “the new fifth bestie,” effectively occupying a rotating guest chair next to the remaining Besties. (allin.onpodcastai.com) In early 2025, episodes repeatedly follow the pattern of the four hosts plus a prominent guest—e.g., E209 “2025 Predictions with bestie Gavin Baker,” E210 featuring Cyan Banister, and E218 with Joe Lonsdale—each billed as guests while the core Besties remain hosts. (allin.onpodcastai.com) Through 2025 the feed also includes a steady stream of such “friends of the pod and newsmakers” (e.g., Nobel laureate John Martinis, María Corina Machado, Bryan Johnson, Ari Emanuel), again framed as guests while the four Besties remain the named hosts. (podbay.fm) External coverage in outlets like the Wall Street Journal describes All-In in this period as a four-host show that now regularly attracts cabinet secretaries, Nobel winners, and other high-profile figures as guests, consistent with a rotating fifth-seat format rather than a permanent new co-host. (wsj.com)

Given that (a) Sacks did in fact come back as a host in future episodes, and (b) thereafter the show consistently used a rotating guest spot for various recurring friends and newsmakers, Jason’s prediction is best classified as right.

The prediction concerns conditions by the year 2030 (specifically, about 100,000 operational eVTOL “flying cars” in China used as passenger air taxis). The current date is November 30, 2025, so 2030 has not yet occurred. There is no way to verify now whether this 2030 outcome will be met. Therefore, the correctness of the prediction is too early to determine.

Chamath’s claim has two parts: (1) capability — that AI + social media will let ordinary people rapidly digest long bills and tell representatives what they want, and (2) outcome — that this will yield a “much more active” and responsive U.S. government where major bills live or die based on that AI‑enabled feedback.

On the capability side, there is clear evidence this is emerging:

• Elon Musk’s Grok on X has been explicitly pitched as a tool that will summarize “mammoth laws” for citizens before Congress votes on them, directly tying LLMs to social‑media distribution. (westernjournal.com)
• Citizen‑facing apps like Represent provide AI bill summaries, personal impact analysis, and an AI “message assistant” that drafts communications to your representatives and tracks your outreach, exactly the workflow Chamath described. (therepresent.app)
• Professional tools such as Plural, Quorum, and FiscalNote now offer AI bill summaries, topic tagging, and predictive analytics on U.S. legislation, showing widespread adoption of LLMs in legislative analysis (though mostly by lobbyists and organized advocates, not the mass public). (pluralpolicy.com)

On the outcome/governance side, there is not yet strong evidence that this has produced a noticeably more responsive form of U.S. governance driven by AI‑enabled grassroots bill analysis:

• Controversy over the One Big Beautiful Bill Act’s proposed 10‑year ban on state AI regulation did trigger broad opposition and a 99–1 Senate vote to strike the ban, but that mobilization appears driven by state officials, civil‑rights groups, and traditional advocacy networks using letters, open statements, and conventional digital activism, not clearly by masses of citizens newly empowered by LLM bill‑readers. (en.wikipedia.org)
• A 2025 natural‑experiment on Change.org found that adding an in‑platform “write with AI” assistant changed petition text but did not improve petition success, suggesting AI‑assisted messaging does not automatically translate into greater political impact. (arxiv.org)
• Research and commercial deployments show AI mostly boosting the analytic capacity of interest groups and government‑affairs professionals, not yet transforming mass public power over legislative outcomes. (arxiv.org)

Only about 11 months have passed since December 2024, the prediction’s starting point, and it concerns a structural shift that plausibly plays out over several years. Given that the technical preconditions are appearing but clear, measurable changes in overall governmental responsiveness due specifically to this phenomenon have not yet been demonstrated, it is too early to decisively label the prediction right or wrong.

Public information so far does not show xAI actually operating or even having fully installed ~1,000,000 GPUs yet, but the one‑year window from the Dec 20, 2024 episode has not fully elapsed as of Nov 30, 2025, so the prediction can’t be definitively graded.

Key facts:

• xAI’s Colossus started with 100,000 Nvidia H100 GPUs in 2024 and was expanded to about 200,000–230,000 GPUs (mix of H100, H200, and GB200) by mid‑2025.(eonmsk.com)
• A September 2025 survey of available data pegs the operational count at roughly 230,000–250,000 GPUs (150k H100, 50k H200, 30k GB200), with another 110k GB200 chips ‘coming online’ and a projected milestone of 1 million GPUs only by late 2026.(etcjournal.com)
• Colossus 2, a second supercluster, is described in mid‑2025 reports as bringing an initial ~550,000 GB200/GB300 GPUs online “within the coming weeks,” with a roadmap to reach about 1 million GPUs across Colossus 1 and 2 by late 2025 or early 2026—but these pieces are framed as plans and ramp‑up, not as already‑achieved 1M‑GPU deployment.(gearmusk.com)
• Overviews and news stories through late 2025 still talk about Colossus ‘aiming’ or ‘on track’ to scale to 1 million GPUs, rather than confirming that 1 million GPUs are currently installed and operational.(etcjournal.com)

Given (a) evidence that xAI has only a few hundred thousand GPUs online so far, and (b) the fact that the stated ~one‑year horizon from Dec 20, 2024 does not end until around Dec 20, 2025, the outcome of Chamath’s prediction—roughly 100k to ~1M GPUs within that year—is not yet definitively knowable from public sources. It therefore remains inconclusive (too early) rather than clearly right or wrong.

Chamath’s prediction is explicitly about what will happen “over time” with no concrete date or horizon attached, so it is a long‑run structural forecast rather than a 1–2 year call.

As of November 30, 2025:

• Estimates of global enterprise software spending and broader “software and software-related” spending (including IT services, cloud, etc.) have not fallen by anything close to 10x. If anything, major analyst firms continue to project continued growth in worldwide software and IT services spending through the late 2020s rather than a collapse.
• However, the claim is that AI will eventually drive the cost base down so much that total annual spend compresses from about $5T to ~$500B. Because no specific date (e.g., by 2030/2035) was given, current data can’t falsify a forecast that could be many years or even decades out.

Since:

1. The predicted 10x shrinkage clearly has not occurred yet, but
2. The prediction does not specify a time frame, and there is no widely accepted evidence that such a long‑run outcome is impossible,

the status of the prediction as of now must be marked as inconclusive (too early) rather than right or wrong.

The prediction’s timeframe has not elapsed yet. Friedberg was effectively saying that by roughly seven years after December 20, 2024 (around 2031–2032) it will be standard practice for organizations to use AI systems that can automatically design and generate production software (including security/compliance logic) from natural‑language instructions and analysis of existing software.

As of November 30, 2025:

• Advanced AI coding agents such as Devin can already plan, write, test, and refactor code autonomously, integrate with tools like Slack/Jira/Linear, learn from existing codebases, and generate production pull requests for real companies, including large refactors at a bank‑like fintech (Nubank).(devin.ai)
• These tools are still in early‑access / pilot‑style deployments and are framed as powerful assistants or teammates whose work must be reviewed before deployment, not as universally adopted, fully trusted, compliance‑aware generators of end‑to‑end production systems across organizations.(newbits.ai)

However, the claim is about what will be standard practice by around 2031–2032, and we are only in 2025. Even though we can say the prediction is not yet fully realized, we cannot judge its final correctness because the deadline is several years in the future.

Therefore, the appropriate status is: it’s too early to tell whether the prediction will ultimately be right or wrong.

As of November 30, 2025, Jason’s prediction is about a 10‑year horizon (from December 2024 to roughly the mid‑2030s). That means we can only falsify it early if the end state he describes has already been reached significantly sooner than a decade.

Jason’s claim, normalized: it will take roughly another decade before fully AI‑generated software systems (i.e., systems whose codebases are largely or entirely produced by AI rather than human authors) are sustainably deployable at high quality in heavily regulated, mission‑critical environments (e.g., medical devices, aviation, critical financial infrastructure, safety‑critical industrial control, etc.).

Current evidence in 2024–2025 points the opposite way: AI‑generated code is widely used but remains error‑prone and security‑risky, and it is typically deployed with substantial human oversight and testing, not as fully autonomous end‑to‑end system generation.

Key signals:

• A Veracode study found that about 45% of AI‑generated code contained security flaws, with no clear security advantage in newer or larger models, underscoring that apparently production‑ready AI code often hides serious vulnerabilities. (techradar.com)
• An Aikido report found that AI‑generated code now accounts for nearly a quarter of production code and is already responsible for one in five major security breaches; 69% of developers and security professionals have seen serious vulnerabilities in AI‑written code, and only a small minority believe secure code can be produced without human oversight. (itpro.com)
• Large‑scale academic analyses of AI‑generated code on GitHub and controlled benchmarks show that while much of it is functional, it systematically introduces bugs and security vulnerabilities, including hard‑coded secrets and path traversal issues, and is overall more prone to high‑risk security problems than human‑written code. (arxiv.org)
• The emerging practice of "vibe coding" (heavy reliance on LLMs to write code from natural language) has been associated with concrete security failures; for example, one vibe‑coding startup’s generated apps had widespread vulnerabilities exposing personal data. Even advocates note that generative systems struggle with complex, multi‑file, safety‑critical software and pose serious maintainability and debugging challenges. (en.wikipedia.org)
• Regulators and insurers are tightening around AI rather than treating it as ready for unsupervised mission‑critical deployment: major insurers are moving to exclude or limit AI‑related risks due to the opaque and unpredictable behavior of large models, especially in high‑stakes contexts. (ft.com)

Taken together, these sources indicate that in 2024–2025, AI is a powerful assistant within software engineering pipelines, but not yet trusted to autonomously generate entire safety‑ or mission‑critical systems in regulated sectors without intensive human control, verification, and traditional assurance processes.

However, Jason’s claim is about the future pace of progress through the mid‑2030s, not about the exact state of 2025. We are only ~11 months into a 10‑year window. There is no evidence that his timeline is already violated (i.e., that we have already reached broadly accepted, fully AI‑generated, high‑quality mission‑critical systems in regulated environments), but it is also far too early to know whether his decade‑scale forecast will ultimately prove accurate.

Because the forecast’s target date lies roughly 9–10 years in the future and current evidence neither confirms nor contradicts the eventual outcome, the correct status as of November 30, 2025 is:

inconclusive (too early to tell), but not falsified so far.

As of November 30, 2025, Google has not launched a clearly defined ad-supported version of the Gemini AI assistant with a corresponding “pay to remove ads” option.

Evidence:

• On Alphabet’s Q4 2024 and early‑2025 earnings calls, Sundar Pichai repeatedly stated that Gemini does not currently have ads and that monetisation is focused on a free tier plus paid subscriptions, while noting they have “very good ideas” for future native ad concepts in Gemini and will likely explore ads later, not in 2025. (theverge.com)
• Multiple reports summarising those calls reiterate that Gemini is presently monetised through free access and paid tiers, not an ad‑supported vs ad‑free split. (news.aibase.com)
• A November 3, 2025 report says Google is experimenting with ads in AI experiences such as “AI Mode” and the Gemini assistant, and that ads are expected to appear in the free versions of Gemini, but also states that a full roll‑out will not happen in 2025—indicating plans and early tests, not a launched, stable business model. (voice.lapaas.com)
• Current public descriptions of Gemini pricing (as of mid‑2025) list a free tier, a $20/month Gemini/AI Pro tier, and a higher‑priced AI Ultra tier, with no mention of an explicit ad‑supported tier where users pay specifically to remove ads from Gemini chats. (tech.yahoo.com)
• Google has introduced ads into AI-powered search overviews / “AI Mode” in Search (which use Gemini technology), but these are framed as search ads inside AI summaries, not as part of a Gemini assistant freemium ad/no‑ad subscription structure. (impressiondigital.com)

Friedberg’s prediction is about what will eventually happen, with no specific time horizon. The available evidence shows Google moving in the predicted direction (planning and testing native ads, hinting at user “options” similar to YouTube), but not yet delivering a full ad-supported Gemini assistant with a paid no‑ads option. Since the relevant future event could still occur after 2025, it’s too early to label the prediction as either right or wrong.

As of November 30, 2025, there is strong evidence of rapid price compression and intense competition, but not enough time or data to say Chamath’s long‑run structural claim has clearly succeeded or failed.

What the prediction requires

• He isn’t just saying prices will fall; he’s saying that for large enterprises, both closed and open‑source model providers will eventually price API usage at (effectively) compute cost, with near‑zero margin per token.
• The phrase “over time” gives no concrete horizon (e.g., 2 years vs. 10 years), so it’s a long‑term industry-structure prediction.

Where pricing actually is in late 2025

1. Closed‑source leaders still charge non‑trivial per‑token prices:

   • OpenAI’s public pricing for major models (e.g., o3, o4‑mini, gpt‑4o‑mini) remains in the roughly $0.15–$20 per million tokens range, depending on model and tier, well above zero and with clearly positive gross margins. (platform.openai.com)
   • Anthropic Claude 4.x models (Opus, Sonnet, Haiku 4.5) list at about $0.80–$15 per million input tokens and $4–$75 per million output tokens, again indicating substantial markups over raw compute. (claudelog.com)
   • Google’s Gemini API charges around $0.15 per million input tokens (with additional output pricing), not zero, even if cheaper than some rivals. (ai.google.dev)

2. Gross margins show providers are still earning more than bare compute cost:

   • Industry analyses estimate OpenAI’s model APIs running at double‑digit to ~50% gross margins, and Anthropic in a similar ~50–60% range—far from “near‑zero” margin over compute. (getmonetizely.com)
   • That implies enterprises are still paying meaningfully above underlying GPU/TPU costs, even after volume discounts.

3. Open‑source and low‑cost competitors are putting pressure on prices:

   • DeepSeek offers extremely low API prices (on the order of $0.55 per million input tokens and $2.19 per million output tokens as of early 2025) and repeatedly cuts prices 50–75% in “price war” moves, explicitly leveraging open‑source releases to undercut Western providers. (en.wikipedia.org)
   • These moves have triggered price cuts from OpenAI and Google and pushed the whole market towards cheaper tiers and smaller, more efficient models, aligning with the direction of Chamath’s thesis (competition and open‑source driving prices down).

Why this is still "too early" rather than clearly wrong

• The current state: prices are falling, there’s a visible race to the bottom on lower‑tier models, and open‑source/China‑based providers have made very cheap inference widely available. That supports the trend he described.
• But the end state he claimed—per‑token enterprise pricing effectively at compute cost across both closed and open providers—has not been reached: list prices and margin estimates show meaningful markups today, and none of the major Western vendors publicly commit to “at‑cost” enterprise tokens. (platform.openai.com)
• Because he did not specify a timeframe, and we’re only ~11 months past the December 20, 2024 episode, the market could still evolve toward his asymptotic scenario (e.g., if vendors shift to primarily charging for dedicated capacity/SLAs while metered tokens trend toward cost).

Given the open‑ended timeline and the fact that pricing is clearly moving downward but is not yet at near‑zero margin, the fairest assessment as of late 2025 is **“inconclusive (too early)” rather than definitively right or wrong.

The prediction is defined on a three‑year horizon: Jason says that by around three years after the episode date (20 Dec 2024), i.e., by late 2027, OpenAI will no longer be a top‑two AI model provider and will instead be roughly 3rd–5th in market position, behind Google Gemini, Meta, and xAI.

As of "today" in this context (30 Nov 2025), fewer than three years have elapsed since the episode date, and we have not yet reached late 2027. Market share and competitive standings can still change substantially between now and then. Therefore, it is too early to determine whether this time‑bound prediction will ultimately be correct or incorrect.

Because the evaluation date precedes the prediction’s deadline, the correctness of the prediction cannot yet be assessed.