Grok 4 Wows, The Bitter Lesson, Elon's Third Party, AI Browsers, SCOTUS backs POTUS on RIFs

As of November 30, 2025, xAI’s Colossus supercomputer (used to train Grok) has not yet reached anything close to 1,000,000 GPUs in operation, but there are multiple public plans and ongoing build‑outs aiming for that scale.

1. Current scale (well below 1M GPUs): Recent coverage summarizing Elon Musk’s statements indicates that xAI currently operates on the order of ~200,000–230,000 GPUs (H100/H200 plus roughly 30,000 Blackwell/GB200 units) for Grok training, far below 1,000,000. (techradar.com) Critical analyses likewise argue that, due to power limits at the Memphis site, only a fraction of the installed GPUs can be powered simultaneously, suggesting effective capacity is even lower in practice. (bluemoonofshanghai.com)

2. Documented plans to reach ~1M GPUs: Well before and after the podcast date, Musk and xAI partners publicly described plans to scale Colossus to around or over one million GPUs. Tom’s Hardware, citing the Financial Times, reports that xAI intends to expand Colossus from over 100,000 H100s to more than one million GPUs, making it one of the most powerful supercomputers in the world. (tomshardware.com) Supermicro has also publicly stated it is supporting xAI’s “10‑fold expansion … with over 1 million GPUs” at the Memphis facility. (tomshardware.com) Local economic‑development announcements and follow‑on reporting describe an expansion that will “incorporate a minimum of one million GPUs,” with project completion targeted in 2026. (eonmsk.com)

3. No evidence of cancellation, but also no 1M‑GPU cluster yet: Energy‑infrastructure reports and power‑plant acquisitions explicitly frame the one‑million‑GPU configuration as a future goal, discussing the power needs and construction work required for such a cluster, not an already‑achieved reality. (finance.yahoo.com) None of the major technical or business sources surveyed claim that xAI has actually deployed and brought online ~1,000,000 GPUs as of late 2025.

Because Chamath’s prediction is that xAI’s Grok cluster will be scaled “up to a million” over time, and the available evidence shows:

• substantial progress beyond the initial 100k/250k plans, yet
• the system is still far short of one million GPUs, and
• the publicly stated timeline to reach ~1M extends into 2026 and beyond,

it is too early to say whether the prediction ultimately proves true or false. The outcome depends on whether xAI actually completes the planned scale‑up in the coming years. Therefore, the appropriate classification for now is “inconclusive (too early).”

Chamath is paraphrasing Elon Musk’s claim on the All-In podcast that later Grok models would move away from “traditional datasets that exist in the wild” and instead be trained via agents generating synthetic data.

Public evidence about Grok’s evolution shows that xAI has not stopped using conventional human-created corpora:

• Regulators in Europe opened an investigation into X’s use of public posts from EU users to train Grok’s LLMs, explicitly describing Grok as trained on large scraped online datasets (articles, blog posts, and social‑media content). This is classic “in‑the‑wild” human data, and there has been no indication that later Grok versions abandoned such sources globally; the reported remedy was limited to EU user data. (apnews.com)
• Reporting on Grok 3 emphasizes synthetic data as an important component, but describes it as in addition to a larger, more diverse dataset rather than a complete replacement of real‑world text. The model is portrayed as mixing synthetic data with real data to improve reasoning and reduce hallucinations, not as being trained solely on synthetic corpora. (rdworldonline.com)
• Musk later said xAI would retrain Grok on a “revised base of human knowledge,” i.e., a re‑edited corpus meant to remove “garbage” and add missing information. That still implies reliance on large human‑authored text collections, just more curated, rather than a pure agent‑generated synthetic dataset. (businessinsider.com)
• Coverage of Grok 5’s planned training states that it will incorporate real‑time data from the X platform to improve relevance and accuracy—a direct continuation of using live, user‑generated social‑media content as part of the training or fine‑tuning pipeline. (grokmag.com)
• xAI’s own materials for Grok 4 and Grok 4.1 talk about large‑scale reinforcement learning and frontier “agentic reasoning” reward models, but they do not claim that the underlying pretraining data has stopped coming from internet and document corpora, and no independent technical source reports such a drastic shift. (x.ai)

Musk has indeed argued that human‑generated data is becoming “exhausted” and that future progress will lean more on synthetic data, which matches the spirit of what Chamath repeated. (theguardian.com) However, available reporting shows that as of late 2025, Grok’s successor models (Grok 4.1/4.1 Fast and the in‑training Grok 5) still depend significantly on conventional human‑authored text from the web, legal documents, and especially X posts, with synthetic data layered on top rather than used exclusively.

Because the prediction was categorical (“not going to be trained on any traditional dataset”), and the ongoing use of real‑world internet and social‑media data is well documented, the prediction has not come true based on what is publicly known, even allowing for some uncertainty about xAI’s proprietary datasets.

Public information shows Elon Musk and xAI clearly intend to lean heavily on synthetic, agent-generated data for Grok, but there is no verifiable evidence that Grok’s training has already shifted to being primarily driven by such data instead of human-generated datasets.

Key points:

• In the All-In discussion summarized by multiple outlets, Chamath recounts Musk saying that future Grok models won’t train on traditional datasets and will instead rely on AI agents creating synthetic data from scratch to drive all the training. This matches the prediction’s wording, but it is a description of Musk’s plan, not confirmation of what xAI has actually done so far. (okx.com)
• Musk has publicly claimed that AI developers have “exhausted” the available pool of human knowledge and that the “only way” forward is synthetic data, where models write and grade their own content. This supports a strategic shift toward synthetic data, but not the specific claim that it already dominates Grok’s training mix. (theguardian.com)
• Some secondary reports state that Grok 3/3.5 were trained extensively or even “primarily” on synthetic datasets, and that synthetic data is central to Musk’s strategy. However, these are not official technical disclosures from xAI, and they still frame synthetic data as augmenting or improving on human data rather than completely replacing it. (linkedin.com)
• Musk, in a later All-In appearance, describes upcoming Grok models as starting from conventional sources (Wikipedia, books, websites) and then using synthetic methods to clean up and rewrite that information—indicating that human-generated corpora remain a core substrate, with synthetic “corrections” layered on top, not a purely agent-self-play data regime. (m.economictimes.com)
• xAI’s official Grok-4 model documentation and public materials do not disclose the proportion of synthetic vs. human data in training, and independent reporting still highlights large human “AI tutor” teams and RLHF-style pipelines, implying substantial ongoing reliance on human-labeled or human-generated data. (docs.x.ai)

Because xAI has not released transparent training-data breakdowns and the available evidence is a mix of aspirational statements, commentary, and partial reporting, we cannot reliably determine whether Grok’s future training regime has in fact shifted to being primarily driven by agent-generated synthetic data rather than human datasets. The claim could turn out to be correct in the longer term, but as of Nov 30, 2025, its truth value is not knowable from public sources, so the prediction’s status is best judged as ambiguous rather than clearly right or wrong.

Evidence since July 2025 shows Elon Musk’s “America/American Party” failed to become a viable third party and was quickly walked back.

• Musk announced the America Party on X around July 5, 2025, positioning it as a new third party aimed at winning a handful of House and Senate seats. However, early reports noted that it had no formal FEC registration, structure, or candidates, beyond social‑media posts and vague plans. (brb.yahoo.net)
• By August 9, 2025, Business Insider reported that Musk had not followed through on any of the concrete steps needed to build a party (ballot access, organization, etc.), concluding that the America Party was “nowhere to be seen” a month after the announcement. (businessinsider.com)
• On August 20, 2025, Reuters reported that Musk had “backed away” from plans to launch the America Party and was refocusing on his companies and maintaining ties with Republican leaders instead of building a separate party. (reuters.com) Reporting in other outlets the same week similarly described Musk as shelving or abandoning the project for strategic reasons. (nypost.com)
• Encyclopedic coverage now characterizes the American/America Party of 2025 as short‑lived or merely an announced/abandoned project, noting that by late August Musk no longer intended to found or proceed with the party. (en.wikipedia.org)

By November 30, 2025, Musk’s third‑party effort has produced no ballot line, no candidates, no elected officials, and no sustained organization, and has effectively been shelved. Under any normal definition of a “viable or successful third party,” this means the “third party stuff” did, in fact, not work, so Sacks’s prediction is best classified as right.

As of November 30, 2025, Elon Musk’s America/American Party has been announced and is organizing but has not yet contested its first federal election cycle.

– Musk launched the “America Party” (often described interchangeably as the American Party) in early July 2025 and has explicitly said it will target a small number of House and Senate seats in the 2026 midterm elections.(upi.com)
– Reporting and analysis consistently frame the party’s electoral impact as something that may occur in the 2026 midterms and possibly the 2028 cycle; there is no reporting of completed House or Senate races actually won or lost by America Party candidates yet.(financialexpress.com)
– The 2026 U.S. House and Senate elections are scheduled for November 3, 2026 and are listed as ongoing/future, meaning no results exist as of late 2025.(en.wikipedia.org)
– Searches for 2025 special elections and FEC activity related to Musk’s party discuss organization, filings, and GOP concern, but not any concluded federal races with America Party nominees, let alone victories.(timesofindia.indiatimes.com)

Because the party’s initial election cycles (starting with 2026) have not yet taken place, we cannot say whether Musk‑backed American Party candidates will win “a few House races” and zero Senate races. There is simply insufficient outcome data so far, so the prediction cannot yet be judged.

As of November 30, 2025, the U.S. Senate’s legislative filibuster rule (the 60‑vote cloture requirement for most legislation) is still in place and has not been abolished. However, Chamath’s prediction is explicitly open‑ended (“just a matter of time,” “on borrowed time”), with no time horizon given. Because he’s predicting an eventual change at some unspecified point in the future, we cannot yet say whether it will ultimately prove right or wrong; we can only say that it has not come true yet.

Given the lack of a deadline or concrete timeframe, the appropriate status is that it’s too early to evaluate the prediction’s correctness, even though the filibuster still exists as of now. Therefore, the prediction is marked as inconclusive rather than right or wrong.

As of November 30, 2025, the U.S. Senate has not eliminated the legislative filibuster; the 60-vote cloture rule for most legislation still applies. However, Chamath’s prediction is explicitly about “some future point” with no time bound, so we cannot yet judge it as right or wrong.

Key facts:

• The Senate nuclear option has previously been used only for nominations: Democrats in 2013 ended the filibuster for most executive and judicial nominees (not Supreme Court), and Republicans in 2017 extended this to Supreme Court nominees. The legislative filibuster, requiring 60 votes for cloture on most bills, remained in place after those changes.
• Searches of 2024–2025 news and Congressional records show no rule change abolishing or effectively neutralizing the filibuster for ordinary legislation; high-profile fights (e.g., over voting rights, abortion, and budget issues) continued to acknowledge the 60-vote requirement.

Because Chamath predicted that eventually a Senate majority leader will “steamroll” the filibuster, and because that event has not yet occurred but also could still occur in the future, the correct status as of now is “inconclusive (too early)”, not right or wrong.

On July 8, 2025, the Supreme Court granted the government’s emergency application to stay a preliminary injunction against President Trump’s broad workforce‑reduction Executive Order 14210 and an OMB/OPM memorandum. That 8–1 order allowed agencies to plan and begin large‑scale reductions in force, while explicitly stating the Court expressed “no view on the legality of any Agency RIF and Reorganization Plan” produced under the order. (scotusblog.com)

On July 14, 2025, in McMahon v. New York (No. 24A1203), the Court considered a specific large‑scale reduction in force at the Department of Education, which sought to eliminate roughly half the department’s workforce pursuant to Trump’s education‑reorganization executive order (and related RIF directives). The Court granted a stay of the district court’s injunction, thereby allowing the Education Department’s mass layoffs and restructuring to proceed while litigation continues—i.e., it ruled in favor of allowing implementation of this specific plan. (caselaw.findlaw.com)

Contemporaneous reporting on that July 14 order states that the vote was 6–3, with all three liberal justices (Sotomayor, Kagan, Jackson) in dissent. This is a narrower margin than the earlier 8–1 ruling on the general RIF planning order. (upi.com)

Taken together, these events match Sacks’s prediction: once a concrete, large‑scale RIF plan under Trump’s workforce‑reduction program reached the Supreme Court, the Court again allowed it to move forward, and the vote was indeed closer than the original 8–1 decision.

Elon Musk did in fact announce and brand a new political organization as the America/American Party in early July 2025, after running a large poll on X and declaring, “Today, the America Party is formed to give you back your freedom.”(investing.com) Subsequent reporting and polling confirm that the party exists as a named political project, although its formal structure and filings have been halting and limited.(axios.com)

However, Jason Calacanis’s prediction is about a private reference class: roughly half of his “top 10–20 close friends in his ‘friend list’” supposedly joining or publicly supporting that party. There is no public, authoritative list of who Jason considers to be in this specific inner-circle “friend list,” and media coverage does not enumerate such a group in a way that can be matched against America Party supporters.

Press coverage of the America Party names a handful of prominent backers or participants—e.g., Vaibhav Taneja as party treasurer and encouragement/support from figures like Mark Cuban and Anthony Scaramucci(timesofindia.indiatimes.com)—but these sources do not identify them as members of Jason Calacanis’s own top 10–20 friends. Other coverage only quotes Jason offering strategic advice about how Elon should use America PAC; it does not indicate that Calacanis himself or his close friends have formally joined or endorsed the party in a way that could be tallied against his “friend list.”(ktik.com)

Because:

1. The composition of Jason Calacanis’s “top 10–20 close friends in his friend list” is not publicly specified or verifiable, and
2. There is no comprehensive, public accounting of which of those specific individuals have “joined or publicly supported” Musk’s America Party,

there is no reliable way to quantify the fraction of that group that has done so, now or in the future. Even though the party has been announced and discussed, the prediction cannot be measured against available evidence.

Therefore, the correct classification is "ambiguous": it cannot be determined from public information whether the prediction proved right or wrong.