E111: Microsoft to invest $10B in OpenAI, generative AI hype, America's over-classification problem

Public information up to November 30, 2025 shows that Google has not fully open‑sourced its main large AI models in the way Chamath described.

• Gemini, Google’s flagship LLM family (used in Search), is proprietary. Gemini (including Gemini 3) remains closed‑source; Google exposes it through paid services (Search, Gemini app, APIs, Vertex AI), and the model weights are not released under an open‑source licence.
  • Wikipedia and other references list Gemini’s license as proprietary and describe it as a commercial product integrated into Google Search and other services, not an open‑source model. (en.wikipedia.org)
• Google did release open‑weights models (Gemma), but they are explicitly not open source. In February 2024 Google introduced Gemma, smaller LLMs whose weights are downloadable and free to use (including commercially), and later expanded the family with larger variants such as Gemma 2 (27B) and Gemma 3. (arstechnica.com) However, multiple reports and even Google’s own briefings emphasize that Gemma is an “open” or “open‑weights” model family, not open source in the traditional/OSI sense. TechCrunch and Forbes both stress that while the weights are widely available, the models are governed by custom terms and are “open (but not open source)” / “open weight, not open source.” (techcrunch.com)
• The specific strategic move Chamath predicted has not occurred. His claim was that "Google will open source their models" and "scorch the earth" by making them widely available and as free as possible to reinforce the value of Google Search. In reality:
  • Google’s core search models (Gemini variants) remain closed and proprietary. (en.wikipedia.org)
  • The "open" initiative (Gemma) targets developers and the broader ecosystem and is not clearly positioned as a deliberate, search‑defensive "scorch the earth" strategy, nor does it involve fully open‑sourcing Google’s main LLMs.

Because Google has not open‑sourced its primary large AI models powering Search—only released smaller, open‑weights models under non‑OSI licences—the prediction, taken at face value, has not come true, even if parts of the ecosystem (like Gemma) rhyme with the general idea of wider, cheaper access.

The core mechanism of Chamath’s prediction—that the swing from surplus to deficit would leave Gavin Newsom unable to significantly expand mental‑health and homelessness initiatives—did not hold up, even though the deficit itself was real.

After a near-$100 billion surplus, Newsom’s January 2023 budget indeed projected about a $22.5 billion gap, a dramatic swing. But he explicitly argued the shortfall would not stop major ambitions in areas including homelessness and housing, and his 2023–24 plan continued funding for housing and homelessness programs such as Homekey and encampment resolution grants. (edpolicyinca.org)

Subsequent budgets under continued fiscal stress preserved and expanded homelessness and behavioral‑health spending. The 2024–25 proposal closed a roughly $37.9 billion shortfall while maintaining a multi‑year $15.3 billion homelessness plan, including $400 million for encampment resolution and $1 billion for Homeless Housing, Assistance and Prevention (HHAP) grants—levels described as more than ever before in state history. (gov.ca.gov) In parallel, Newsom put a major behavioral‑health and housing overhaul on the March 2024 ballot (Proposition 1): a $6.4 billion bond plus Mental Health Services Act reforms, intended to build thousands of treatment beds and supportive housing units targeted heavily at people experiencing homelessness. Voters narrowly approved it, and state guidance describes 6,800 new behavioral‑health beds, 26,700 outpatient treatment slots, and 4,350 permanent supportive housing units (over half reserved for veterans). (gov.ca.gov)

Implementation since then confirms this is not a case of being “unable to do any of this stuff.” In May 2025, Newsom announced $3.3 billion in Proposition 1 grants to create more than 5,000 residential treatment beds and over 21,800 outpatient slots for mental‑health and substance‑use care, with a specific focus on people experiencing homelessness. (gov.ca.gov) In February 2025 he also released $920 million in additional state homelessness funding tied to new accountability measures for local governments. (gov.ca.gov) These are large, new (or newly scaled) funding streams, not the absence of expansion.

On outcomes, California’s homelessness problem did get somewhat worse in the short term, but not in the runaway way implied by a state unable to invest. HUD data show California’s homeless population at about 181,399 people in 2023 and roughly 187,084 in January 2024—a roughly 3% statewide increase. (spesmea.org) State and federal summaries characterize this as continued growth but at a slower rate than the national average, with unsheltered homelessness in California rising less than 1% while increasing more sharply elsewhere. (hcd.ca.gov) Preliminary 2025 information is mixed but suggests some local improvement: Newsom’s office highlighted that early 2025 point‑in‑time data from several large communities show declines, and Los Angeles County reported a 4% drop in homelessness in 2025—its second consecutive annual decline—along with a larger fall in unsheltered and chronic homelessness in the City of L.A. (gov.ca.gov)

Netting this out: (1) the specific fiscal claim—that the deficit would prevent Newsom from materially expanding mental‑health and homelessness initiatives—has been clearly falsified by the passage and implementation of large new programs like Proposition 1 and continued multi‑billion‑dollar state homelessness funding; (2) the outcome claim—that homelessness would likely get worse in the near term—was only partially borne out, with a modest statewide increase into 2024 followed by signs of stabilization or improvement in several major jurisdictions by 2025. Because the central budget‑constraint premise is wrong and the worsening of homelessness has been limited rather than driven by an inability to fund initiatives, the overall prediction is best judged wrong.

Jason’s prediction was that, absent a fundamental shift in how society treats fentanyl‑type drugs, problems like addiction, overdoses, and visible street use would “just get worse” from the 2023 baseline.

Empirically, U.S. overdose deaths have not continued to worsen since 2023. CDC data compiled by researchers show overdose deaths peaking around 110–111k in 2022, then falling to about 108.6k in 2023 and about 81.7k in 2024, with roughly 76.5k deaths in the 12 months ending April 2025—a substantial and continuing decline rather than further worsening. (en.wikipedia.org) Independent analyses report that 2024 saw about an 27% drop in total overdose deaths versus 2023—the steepest one‑year decline ever recorded—and roughly a 37% reduction in deaths involving synthetic opioids like fentanyl. (reuters.com) These are major improvements relative to the 2023 baseline Jason was referencing.

Local data from hard‑hit cities show similar patterns: San Francisco hit a record high in overdose deaths in 2023, driven largely by fentanyl, but 2024 saw roughly a 20–22% decline in overdose fatalities and the lowest monthly tolls in several years, even though levels remain elevated and there has been some rebound in early 2025. (voz.us) That again contradicts a claim that things would unambiguously “just get worse” from that point.

On the policy/recognition side, the U.S. has in fact increasingly singled out fentanyl and similar synthetics as a distinct, unusually dangerous threat. The Biden administration’s National Drug Control Strategy and subsequent fact sheets emphasize targeted action on “illicit fentanyl supply chains,” large budget increases for fentanyl‑focused enforcement, and national campaigns like Real Deal on Fentanyl, plus historic expansion of naloxone access and other harm‑reduction and treatment measures. (bidenwhitehouse.archives.gov) Legislatures have also advanced fentanyl‑specific criminal penalties (e.g., federal bills like the Felony Murder for Deadly Fentanyl Distribution Act) and broader tough‑on‑drug measures such as California’s Proposition 36, along with local crackdowns on open‑air fentanyl markets in cities like San Francisco. (en.wikipedia.org)

Social harms from synthetic opioids are still severe—overdose deaths remain far above pre‑2019 levels, street drug scenes persist in many cities, and new ultra‑potent drugs (e.g., nitazenes, carfentanil) are emerging. (en.wikipedia.org) But Jason’s core forecast was that there was “no path” and that, from the 2023 baseline, these problems would continue to worsen over time unless society fundamentally reclassified and policed fentanyl‑type drugs differently. Instead, overdose mortality and some related indicators have measurably improved over 2023–2025 while recognition and targeted responses have intensified.

Because the key observable outcome he pointed to—worsening overdose and visible drug problems from the 2023 baseline—has moved in the opposite direction nationally (and in major hotspot cities), the prediction as stated is best judged wrong.

As of November 30, 2025, neither the legal system nor the market has clearly settled in the way Jason predicted, but there are meaningful moves in that direction, so the prediction can’t be cleanly called right or wrong.

Legal requirements / permission for data owners

• In the U.S., there is still no specific statute that requires generative AI providers to get permission from all copyright holders whose works are used in training. Courts have issued early rulings treating some training on copyrighted books as fair use (e.g., Meta’s Llama case and Anthropic’s Claude case), confirming that permission is not categorically required, though the law remains unsettled and many lawsuits are ongoing. (theguardian.com)
• In the EU, the AI Act now imposes transparency and copyright-compliance duties on providers of general‑purpose AI models. They must publish a summary of training data and respect copyright, including honoring rights‑holders’ opt‑out from text and data mining; if a rightsholder opts out, authorization is required. But the regime is based on exceptions plus opt‑outs, not a blanket “permission for everything” rule. (hoganlovells.com)
• In China, interim measures for generative AI require use of “legitimate data sources,” non‑infringement of IP, and consent for personal data, again signaling stricter standards but not a globally uniform mandate that all training data be licensed. (staticpacific.blob.core.windows.net)
• In the U.S., the proposed Generative AI Copyright Disclosure Act would require companies to disclose which copyrighted works (with URLs) they used in training, but it explicitly does not ban use of copyrighted works for training and, as of late 2025, remains only proposed legislation, not enacted law. (en.wikipedia.org)

Market pressure, licensing deals, and citations

• Major AI vendors have voluntarily struck content‑licensing deals with publishers (e.g., OpenAI with the Associated Press, Axel Springer, and Condé Nast; Perplexity with Le Monde), which allow use of that content for training and display, including attribution and links in answers. This is clear evidence of market pressure toward licensing and attribution, but it is selective and far from universal. (cnbc.com)
• At the same time, many high‑scale models still appear to rely heavily on large web scrapes and fair‑use/text‑and‑data‑mining theories rather than comprehensive opt‑in licensing, as reflected in ongoing copyright lawsuits by news and reference publishers against OpenAI and Perplexity. (theverge.com)

Citation / linking behavior in products

• Some generative AI services, especially those positioned as search tools, now routinely provide citations and links. OpenAI’s ChatGPT Search integrates a search engine that “generates responses, including citations to external websites,” and Deep Research is built explicitly to create cited reports. (en.wikipedia.org)
• Competing systems like Perplexity and Bing Copilot similarly foreground linked sources, but empirical work on “attribution gaps” shows that web‑enabled LLMs frequently fail to credit many of the pages they read, and many responses contain few or no clickable citations. This underscores that citation practices are product‑design choices, not mandatory legal obligations. (arxiv.org)
• There is, as of 2025, no law in the U.S. or EU that generically requires chatbots like ChatGPT to attach per‑answer citations to their training data or even to all web sources consulted, though EU rules do require public summaries of training data for large models and respect for copyright and opt‑outs. (hoganlovells.com)

Why this is ambiguous rather than right or wrong

• Jason’s claim had two parts: (1) services like ChatGPT would be required to use citations and links, and (2) they would be required to get permission from data owners to operate at scale. As of late 2025:
  • Citations and links are common in some product modes (especially search) but not universally mandatory across all generative AI use cases.
  • Permission/licensing is becoming more common for certain high‑value content and is reinforced by EU and Chinese rules, but large‑scale models can and do still operate relying on fair use or text‑and‑data‑mining exceptions without comprehensive opt‑in permission.
• Because his statement is framed as an ultimate end state (“that’s where this is all going to shake out”) with no explicit time horizon, and regulatory and market trends are still evolving—some in the direction he predicts, others not—we cannot yet say his forecast has clearly succeeded or clearly failed.

Given this mix of partial convergence (more licensing, more attribution in some products, new transparency rules) and clear gaps (no global obligation for per‑answer citations or universal permissions, plus ongoing reliance on fair‑use/TDM), the fairest classification today is ambiguous rather than decisively right, wrong, or simply “too early” in a time‑bound sense.

Friedberg’s claim is that over the coming years “essentially all” software-dependent business models rooted in traditional information retrieval would have to be rewritten to incorporate AI-driven synthesis.

Evidence through late 2025 shows:

• Widespread but not universal adoption: Surveys find ~72–78% of organizations use AI in at least one function and ~65–71% use generative AI, indicating rapid uptake but not near‑total penetration, and usually in specific functions rather than full business‑model rewrites. (mckinsey.com)
• Enterprise software not yet structurally transformed: Gartner estimates less than 10% of enterprise software was multimodal/GenAI‑infused in 2024, projecting 80% by 2030, which explicitly frames deep AI integration as a transformation still mostly in the future. (gartner.com)
• Core information‑retrieval businesses remain largely intact: In web search, Google still holds ~89–91% share in 2025; AI‑native or AI‑centric search tools (ChatGPT, Perplexity, etc.) represent only about 5–10% of queries, and Google’s business model is still primarily ad‑driven search, albeit with AI overlays. (infront.com) This suggests IR‑based models have been augmented, not yet fundamentally replaced en masse.
• Most companies are still experimenting, not fully restructured: Gartner, BCG, MIT and others report that although many firms are piloting or deploying GenAI, only a small minority achieve strong, measurable business value; many AI projects are siloed, underperforming, or stalled due to integration challenges and legacy systems. (businessinsider.com)

These data show a clear direction of travel toward AI-driven synthesis inside software and workflows, consistent with the spirit of the prediction, but they fall far short (so far) of “essentially all” IR‑based software business models being restructured or replaced. Because Friedberg’s horizon was “over the coming years” and major analysts themselves place the bulk of this structural shift closer to 2030, it is too early to decisively judge the prediction as either right or wrong.

Hence, the status as of November 2025 is inconclusive (too early).

As of late 2025, the prediction that OpenAI‑like foundational capabilities would be widely available from multiple vendors and at least partially commoditized is essentially correct.

1. Multiple vendors with broadly similar foundational capabilities

Since early 2023, many large‑scale, general‑purpose LLMs comparable in kind to OpenAI’s GPT family have been released and productized by different organizations:

• Anthropic: Claude 3 (Opus, Sonnet, Haiku) and later Claude 3.5 models provide GPT‑4‑class reasoning and are widely available via API and in third‑party products.
• Google: Gemini (Ultra, Pro, Flash) and its successors are positioned as general‑purpose foundation models integrated across Google Cloud and consumer products (Search, Workspace, Android, etc.).
• Meta: LLaMA 2 and LLaMA 3 are open‑weight models explicitly released as general foundation models for commercial and research use, widely adopted and fine‑tuned by many companies.
• Mistral and other startups: Mistral’s Mixtral and later models, plus many other open and closed LLMs, provide competitive capabilities and are accessible via standard APIs.

The net effect is that core LLM capabilities (chat, coding assistance, summarization, Q&A, etc.) are no longer exclusive to a single vendor; they are available from several big tech companies and multiple startups, plus open‑weight models that anyone can deploy.

2. Evidence of partial commoditization

“Commoditized” in the context of the quote means: the base capability (a strong general‑purpose LLM) is available from many suppliers with decreasing differentiation and with competition driven partly on price, latency, and deployment form factor rather than unique raw capability.

By 2024–2025, we see features of exactly this:

• Price competition and similar SKUs: Major providers all expose comparable tiers (fast/cheap vs. large/expensive models) and frequently cut prices or introduce cheaper distilled models. This is classic commoditization behavior.
• Standardized interfaces: Many providers follow very similar JSON/chat/completions APIs, and tools like LangChain, LlamaIndex, and various “multi‑provider” SDKs treat models from OpenAI, Anthropic, Google, and open‑weights as largely swappable components.
• Open‑weight alternatives: LLaMA 2/3, Mistral, and other open‑weight models allow enterprises to self‑host competitive models on their own infrastructure or edge devices, further eroding any exclusive moat around simply having a capable base LLM.
• Differentiation shifting up the stack: Most competitive AI products now differentiate with fine‑tuning, retrieval‑augmented generation, proprietary data, UX, integration, and domain‑specific tooling, not just which base model they use. This is strong evidence that the foundational model itself is treated more like a commodity input in many contexts.

At the same time, it is not fully commoditized in a strict economic sense—frontier models (e.g., the very top‑end proprietary models) are still concentrated in a few labs and can have meaningful capability gaps. But the prediction only claimed “eventually going to become commoditized or certainly much more widely available… there’ll be multiple players that offer them” and that the durable advantage from the core models themselves would be reduced, not eliminated.

Given:

• The existence of many broadly similar LLM providers.
• Active price and performance competition.
• Open‑weight models narrowing the gap and enabling self‑hosting.
• Strategic focus shifting away from raw model access toward data, integration, and product.

…it is reasonable to conclude that the prediction has effectively come true in the “partially commoditized, widely available from multiple players” sense that Sacks described.

Chamath’s claim was that, with about a year of intensive associate feedback (“bang on some law model for a year”), a legal-domain model’s precision and recall would be “off the charts” and effectively perfect, suitable for high‑stakes legal use.

Since early 2023, the closest real‑world tests of this hypothesis have been specialized legal LLMs (Harvey, Lexis+ AI, Westlaw AI tools, etc.) that have indeed had heavy use and feedback from thousands of lawyers over well more than a year.

1. Specialized legal tools are far from near‑perfect.

   • A preregistered Stanford study of leading legal research tools (Lexis+ AI, Westlaw AI-Assisted Research, Ask Practical Law AI, GPT‑4) found that even the best system (Lexis+ AI) hallucinated 17–33% of the time and answered only about 65% of queries accurately—nowhere near “perfect” precision/recall. (arxiv.org)
   • A 2025 comparative evaluation likewise found Lexis+ AI to have 58% accuracy and ~20% fabricated responses, with other tools doing worse—again inconsistent with “off‑the‑charts” reliability suitable for unsupervised high‑stakes use. (cambridge.org)

2. Legal‑specialized vendors acknowledge substantial remaining error.

   • Harvey’s own BigLaw Bench results show its assistant model completing about 74% of a lawyer‑quality work product on complex legal tasks and achieving a 68% “Source Score” for correctly sourced answers, with the company explicitly noting “substantial room for improvement,” not perfection. (harvey.ai)
   • In a follow‑up post, Harvey reports a low but non‑zero hallucination rate (~1 in 500 claims, 0.2%) on its internal benchmark—impressive, but still not “perfect,” and limited to its own task distribution. (harvey.ai)

3. The “one‑year of associates” condition is effectively met in practice.

   • Since late 2022, firms such as Allen & Overy (now A&O Shearman) have had thousands of lawyers using Harvey—over 3,500 lawyers making ~40,000 queries just in the early beta—providing exactly the kind of intensive, expert feedback Chamath described. (arstechnica.com)
   • Harvey then collaborated with OpenAI on a custom case‑law model, tested with 10 major law firms, and tuned specifically to reduce hallucinations; lawyers preferred its outputs 97% of the time to baseline GPT‑4, yet even Harvey presents this as a major improvement in reliability and relevance, not as achieving perfection. (openai.com)
     In other words, the industry has roughly executed the scenario Chamath imagined—sustained legal‑expert RL on top models for more than a year—without reaching anything close to universally “perfect” legal performance.

4. Courts, bar associations, and vendors still treat AI as non‑trustworthy for unsupervised high‑stakes work.

   • Stanford and ABA‑linked work on “Hallucinating Law” finds that general LLMs frequently hallucinate in core legal reasoning tasks, with hallucination rates 69–88% overall on certain benchmarks and at least 75% when asked about a court’s holding, concluding that current models are “not yet capable of the nuanced legal reasoning required” for tasks like evaluating precedential relationships. (americanbar.org)
   • Professional guidance and commentary from Thomson Reuters, law firms, and ethics authors consistently stress that AI systems are tools, “not a substitute for a lawyer”, and that human oversight is “critical” because outputs remain fallible and prone to hallucinations. (thomsonreuters.com)
   • Courts across multiple jurisdictions have sanctioned lawyers for filings containing fake AI‑generated citations, with judges explicitly warning that “no reasonably competent attorney should outsource research and writing” to AI without verification. (markets.chroniclejournal.com)

Given that:

• Highly‑invested, domain‑specialized systems have had more than a year of intensive feedback from lawyers,
• Rigorous empirical studies show substantial error and hallucination rates, far from near‑perfect precision/recall, and
• The legal profession and courts still require human review and explicitly warn against relying on these tools in high‑stakes matters,

Chamath’s prediction—that a year of associate‑driven RL would yield effectively perfect, high‑stakes‑grade legal performance—has not materialized. The technology has advanced dramatically, but empirical results and professional practice clearly contradict the level of reliability he forecast.

So the forecast is wrong rather than merely “inconclusive.”

Part of the prediction came true, but in a mixed and short‑lived way, and the 10–20x productivity claim is not clearly borne out.

1. Did “prompt engineer” emerge as a recognized job title and valued skill?
Yes, at least for a time. In 2023–24, major outlets reported “prompt engineer” as a new job title created by the rise of generative AI, explicitly describing it as a specialty focused on crafting prompts for systems like ChatGPT and quoting industry leaders on its importance.(axios.com) High‑profile postings such as Anthropic’s “Prompt Engineer and Librarian” role with a six‑figure salary band, along with salary surveys and 2024–25 guides listing prompt engineer as a discrete role with defined ranges, show that the title was real, niche but visible, and often very well paid.(promptjobs.com) Numerous training and certification providers now market prompt engineering as a core AI skill and “career booster,” reinforcing that it is a recognized and valued capability even when the exact job title is not used.(store.aicerts.ai)

However, by 2024–25 several analyses and labor‑market observers noted that job postings explicitly titled “Prompt Engineer” were relatively rare, with the work increasingly folded into broader roles like ML engineer, AI engineer, or automation architect. An economist at Indeed’s Hiring Lab is quoted saying she rarely sees it as a standalone job title, and multiple 2025 articles argue that prompt engineering has effectively shifted from a distinct job to an embedded skill.(techspot.com) So the title did emerge and gain recognition, but it did not stabilize as a large, enduring standalone profession; instead, the skill became more mainstream while the dedicated role remained small and in some areas is already fading.

2. Are top practitioners widely seen as 10–20x more productive than average knowledge workers using AI tools?
There is a strong narrative that AI “power users” or expert prompt engineers achieve order‑of‑magnitude gains, but this is mostly marketing language and anecdote, not a clearly established labor‑market consensus. Many blogs, courses, and webinars in 2024–25 explicitly claim that mastering prompt engineering or power‑user workflows can give “10x better results” or make people “10x more productive,” often positioning a small group of expert prompters against more casual AI users.(aitoolbook.ai) These claims echo the “10x engineer” trope the predictor referenced.

That said, this 10–20x figure is largely rhetorical. There is little rigorous, large‑scale evidence quantifying prompt engineers as being 10–20 times more productive than average AI‑using knowledge workers, and the messaging tends to lump prompt engineering into broader AI power‑use rather than a narrowly defined “prompt engineer” occupation. At the same time, other analyses emphasize that as models improve and tools help generate prompts automatically, the distinct advantage of specialist prompt engineers is shrinking and their work is being absorbed into more technical or hybrid roles.(techspot.com)

Overall:

• The prediction that prompt engineering would become a recognizable title and valued professional skill within a few years is partly right: the title did appear, gained media and hiring visibility, and commanded high salaries in some postings, while the underlying skill became widely promoted and integrated into many roles.(axios.com)
• The stronger claim—that top prompt engineers would sustainably be seen as 10–20x more productive than average AI‑using knowledge workers—is not clearly established beyond hype and anecdotal marketing, and the standalone role is already being questioned or downplayed in 2025.(techspot.com)

Because some key elements materialized but others are overstated or weakly supported, the fairest overall judgment is ambiguous rather than clearly right or wrong.

The prediction explicitly gives a 3–5 year horizon: Jason says that in “three or 4 or 5 years” it may be possible to run a company like Twitter with 98% fewer employees than its pre-layoff level.

• The podcast was released on January 13, 2023.
• A 3–5 year window from that date runs from January 13, 2026 to January 13, 2028.
• Today is November 30, 2025, which is less than 3 years after the prediction.

Since the earliest deadline (3 years) has not yet arrived, we cannot say whether the prediction turns out right or wrong. There is also no evidence that such extreme automation has already been decisively proven impossible for the 2026–2028 period.

Therefore, the correct classification as of now is: it is too early to tell whether the prediction is right or wrong.

Summary: Parts of Friedberg’s cluster of predictions have largely materialized (AI-written fiction from significant publishers, AI-composed music played by professional orchestras, and AI-generated films with wide online distribution). However, his 3–4 year timeline for mainstream text-to-game tools runs through ~2026–2027, so as of November 2025 it’s too early to fully judge the whole bundle.

1) AI novel from a major publisher in 2023

• In May 2023, Death of an Author by Stephen Marche (as “Aidan Marchine”) was released by Pushkin Industries. Marche has said about 95% of the text is AI‑generated using ChatGPT, Cohere, and Sudowrite; the cover and blurbs are also AI-generated. It has been widely described in major media (e.g., The New York Times) as an early or first “AI novel.” (en.wikipedia.org)
• Pushkin Industries, however, is primarily a prominent podcast/audiobook company, not one of the traditional “Big Five” book publishers; it does publish books but is mainly known as an audio-first media outfit. (en.wikipedia.org)
  Assessment: There was a widely covered, mostly‑AI‑written long-form work released in 2023 by a serious, well‑funded publisher, but whether Pushkin counts as a “major book publisher” in the sense Friedberg probably meant (Penguin Random House, HarperCollins, etc.) is debatable. So this sub‑prediction is partially but ambiguously fulfilled.

2) AI symphony performed by a major symphony orchestra in the near term

• Well before 2023, AI‑composed orchestral music was already being performed by professional orchestras. For example, the AI composer AIVA had works performed by the Avignon Symphonic Orchestra in 2017. (en.wikipedia.org)
• In 2021, the Beethoven Orchestra Bonn premiered Beethoven X: The AI Project, in which AI systems (with human musicologists) completed Beethoven’s unfinished 10th Symphony; the AI-generated Scherzo and Rondo movements were performed in concert and recorded for commercial release. (telekom.com)
• Post‑podcast, additional projects appeared. For instance, Spain’s national RTVE Symphony Orchestra performed two short symphonic works composed by AI at Madrid’s Monumental Theater on November 17, 2023—one rendered exactly as produced by the AI, another lightly rearranged by the conductor. (english.elpais.com)
  Assessment: If you interpret Friedberg as forecasting that within a few years of early 2023, a serious symphony orchestra would perform a piece whose musical material is primarily AI‑composed, that has clearly happened (and in fact had happened even before his prediction). The only quibble is that some early cases (Beethoven X, AIVA) are AI‑assisted/completion rather than entirely original AI symphonies. Overall this sub‑prediction is substantively correct.

3) Primarily AI‑generated screenplay → AI‑generated 3D/CG movie with broad public viewership in the near term

• AI‑written feature scripts:
  • The Diary of Sisyphus (Italy) is described as the first feature‑length film written by an AI (GPT‑Neo). It premiered at festivals in 2023 and was released in Italian cinemas in January 2024, though with human-shot live‑action visuals. (en.wikipedia.org)
  • The Last Screenwriter (2024) is a Swiss feature whose screenplay was written by ChatGPT; the film is live-action and was released online for free distribution after its planned London premiere was canceled. (en.wikipedia.org)
• Primarily AI‑generated movies with wide distribution:
  • DreadClub: Vampire’s Verdict (2024) is an 87‑minute animated feature described as the first fully AI‑generated animated feature film. All visuals, animation, performances, music, sound, and even some editing were produced via AI tools; a single filmmaker orchestrated prompts and structure. It streams worldwide on Amazon Prime Video and other major platforms, making it widely accessible to the general public. (en.wikipedia.org)
  • Industry is now moving toward larger‑budget AI‑heavy productions such as Critterz, a feature‑length animated film being developed with OpenAI’s upcoming models (e.g., GPT‑5, Sora), targeted for a 2026 premiere and explicitly marketed as a proof‑of‑concept for AI‑driven filmmaking. (techradar.com)
    Assessment: By 2024 there is at least one full‑length film whose script and all audiovisual elements are primarily AI‑generated and that is globally streamable, which substantially matches what he envisioned, even though DreadClub is 2D anime‑style rather than explicitly 3D CGI. Given how close the realized projects are to his description (and the trajectory toward larger mainstream releases), this sub‑prediction is basically fulfilled in spirit.

4) Within ~3–4 years (by ~2026–2027), consumer tools that let ordinary users create their own AI‑generated video games via natural‑language specs

• Several tools already point in this direction, but are still limited:
  • Buildbox 4 (2024–25) adds AI features like text‑to‑game translation, where users describe game elements in text and the AI generates assets, scenes, and prototypes. It’s explicitly aimed at non‑programmers who want to build games visually with AI assistance. (en.wikipedia.org)
  • Rosebud AI is a startup whose platform lets users upload textual descriptions of the game they want; the AI interprets this and generates graphics, characters, and game worlds, with creators showing off games made in minutes from prompts. (ru.wikipedia.org)
  • Niche apps like StoryGen let ordinary users create interactive visual‑novel‑style games from text prompts, automatically adding characters, music, and branching choices. (reddit.com)
  • Research and demos like Oasis, an AI‑generated playable Minecraft‑style world run entirely by generative models, show that full AI‑driven, code‑free game worlds are technically possible, though still glitchy and experimental. (en.wikipedia.org)
• However, these tools are still constrained—usually 2D, template‑driven, or requiring significant iterative prompting and manual adjustment. They do not yet consistently deliver the robust, arbitrary, high‑fidelity games most people imagine when they hear “describe any game you want and it appears.”

Assessment: The direction of progress matches Friedberg’s forecast, and early consumer‑facing text‑to‑game systems exist, but his explicit time horizon runs until approximately 2026–2027. As of late 2025, it’s not yet clear whether truly mainstream, widely‑adopted tools that fully match his vision will arrive in that window. So this part is still in progress and not yet judgeable.

Overall verdict:

• One component (AI symphony) is clearly right; another (AI film) is effectively right in spirit; the AI‑novel component is partially/arguably right but depends on how you define a “major” publisher; and the video‑game component’s deadline hasn’t arrived yet. Because at least one major sub‑prediction is still within its time window and cannot yet be definitively graded, the combined prediction set is best labeled inconclusive at this time.

Merrick Garland’s Department of Justice did not drop or scale back the classified-documents case against Donald Trump after the Biden documents discovery.

Key facts:

• On November 18, 2022, Attorney General Merrick Garland appointed Jack Smith as special counsel to oversee investigations involving Donald Trump, including the Mar-a-Lago classified-documents matter. The appointment remained in force after the January 2023 revelations about classified documents found at Biden’s former office and residence.
• On June 9, 2023, a federal grand jury in the Southern District of Florida indicted Trump on 37 counts related to retention of national defense information, obstruction, and false statements arising from his handling of classified documents at Mar-a-Lago.
• On July 27, 2023, a superseding indictment added further charges and an additional defendant, expanding rather than narrowing the case.
• As of late 2025, the Mar-a-Lago documents prosecution has not been dropped; it remains an active federal criminal case, with ongoing pretrial litigation and no indication that DOJ has abandoned or materially scaled back the core classified-documents charges.

Because the DOJ proceeded to charge Trump and has continued to pursue that prosecution, the prediction that Garland would "have to drop the prosecution ... or at least that part" related to the documents at Mar-a-Lago did not come true.