Leitura

Altimeter founder Brad Gerstner appeared on the All-In Podcast, showed SemiAnalysis's data, and stated that the firm's forecast of +43GW in 2027 is too aggressive. He thought only about 25GW would actually get built, close to half of SemiAnalysis's number, based on his "hunch." SemiAnalysis's response was blunt: let's not operate on hunches.

SemiAnalysis's core argument is not a single announcement or verbal commitment, but two trackable leading indicators: quarterly self-build construction starts (measured in megawatts) and datacenter lease contract value. The logic chain is as follows: SemiAnalysis states that given the capital intensity of building multi-hundred-MW datacenters, developers typically only spend 5% to 10% of project capital without a customer. They need a pre-lease — typically a 15- to 20-year contract with a total contract value of $30 to $50 billion per GW — to fund and develop the whole project. Therefore, signed lease value leads debt raises, which in turn leads the actual buildout. Tracking leases means tracking the front end of the construction pipeline.

From H2 2023 through H1 2025, quarterly datacenter leasing remained relatively stable, largely because Microsoft paused its large-scale rental activity — SemiAnalysis claims it called this out before anyone else in December 2024. The turning point came between August and October 2025: Oracle closed on approximately 4.5GW of datacenter IT capacity across three sites. After that, Meta, AWS, Microsoft, Google, Nvidia, AMD, Anthropic, and OpenAI all entered the market for GW-scale leases, with a long tail of neocloud companies signing hundred-megawatt-scale deals. SemiAnalysis reports that just in 2026 year-to-date, it has identified over $1 trillion in lease value that was signed and closed.

What does this over-$1-trillion in leases mean? At $30 to $50 billion total contract value per GW, it corresponds to roughly 20 to 33GW of IT capacity. Combined with projects already under construction but not yet leased, SemiAnalysis believes this is sufficient to support 50GW of new capacity in 2027 — essentially the entire global cumulative datacenter market through 2024 being built in a single year. Its quarterly self-build construction data also shows that US hyperscalers have never launched more datacenter construction than they are now, with the number going parabolic, driven in particular by two hyperscalers.

But this argument has clear boundaries. First, SemiAnalysis's Datacenter Industry Model is a paid subscription product; the public article displays model conclusions rather than the full underlying data, so readers cannot independently verify the completeness of its building-by-building tracking. Second, the article's most critical section — whether datacenters will face oversupply — is locked behind the paywall; the public portion only poses the question without answering it. Third, leases are contractual obligations, not delivered capacity: from signing to energized delivery still requires construction, equipment arrival, and grid connection, and delays at any stage can push 2027's 50GW back. SemiAnalysis itself acknowledges that its model continues to flag developer timelines that appear overly aggressive, including the Oracle New Mexico force majeure site delay in April.

For AI industry investors, the practical use of this analysis is to provide an actionable judgment framework: when the market produces narratives about datacenter buildout slowing or overheating, checking the pace and dollar value of binding, non-cancellable lease signings reflects the real pipeline state better than tracking single announcements or executive statements. Whether the over-$1-trillion in leases ultimately converts into 50GW of delivered capacity depends on execution; but the transmission logic of leases as a leading indicator is a structural variable worth incorporating when assessing AI compute supply prospects.

According to reports from Reuters and the Financial Times, Microsoft has reportedly terminated its data center agreement with UK cloud infrastructure provider Nscale for a site in Loughton, Essex. The facility was originally planned to house approximately 23,000 NVIDIA GB300 chips to support Azure cloud services, but grid operators informed Nscale that power supply would not be available until the 2030s. Microsoft declined to comment and did not explain the exact reasons for the termination.

The Financial Times cited sources familiar with the matter stating that Microsoft had intended to lease the full 50 MW capacity of the site, leaving Nscale to seek new tenants. Nscale told Reuters it remains committed to the Loughton project and is exploring alternative power sources and other expansion sites in the UK, noting that other contracts with Microsoft are unaffected. However, the company has not announced an alternative power solution or a new launch date.

This event underscores that the critical bottleneck for AI infrastructure expansion has shifted from 'acquiring chips' to 'physical grid interconnection.' Despite having capital and chip supply chains, major tech companies face multi-year or decade-long delays due to local grid upgrade cycles, queue times, and regulatory approvals. When power supply cannot meet expected timelines, even announced compute capacities face risks of delay or cancellation. For investors and industry observers, grid interconnection status has become a more urgent limiting factor than chip orders when assessing the feasibility of AI data center projects.

Amazon Web Services (AWS) released a development tool named TorchNeuron on October 19, 2026, designed to allow AI startups to easily migrate their applications to Amazon's custom AI chip, Trainium. This move is seen as a direct challenge to Nvidia's long-standing software advantage, specifically its CUDA ecosystem. CUDA is the critical software infrastructure that enables applications to run on Nvidia GPUs, and its deep integration has made Nvidia the default hardware choice for AI application development.

TorchNeuron is compatible with the open-source framework PyTorch. According to Amazon, approximately 85% of AI apps are written using PyTorch. The tool allows developers to run PyTorch-based applications on Trainium chips by changing just one line of configuration code, without requiring a complete rewrite. Previously, Nvidia spent years ensuring its CUDA software seamlessly supported PyTorch, creating strong user stickiness. While AMD and Apple have offered similar porting options, Amazon emphasizes that migrating to Trainium not only simplifies operations but also grants startups access to Amazon's entire cloud service ecosystem.

Colin Brace, VP of Technology Software at AWS Annapurna Labs, stated that building on AWS means developers do not need to bet on a single-purpose compute provider and scramble to build everything else around it. He noted that the infrastructure needed to go from prototype to production is already there, helping founders scale quickly. Additionally, Amazon claims that because Trainium handles both training and inference tasks and is tightly integrated with cloud infrastructure, it saves startups money through overall workload economics, even if the per-token price is not the lowest. For instance, voice AI company Deepgram reported that after moving to Trainium, they processed 70% more audio per dollar while maintaining production-grade accuracy.

Beyond providing convenient tools, Amazon plans to contribute code to the main PyTorch project to remove parts that tie the framework to specific chips, aiming to make PyTorch more hardware-agnostic for the entire developer community. If successful, this could have deeper ramifications for Nvidia than simple chip competition, as it undermines the lock-in effect that drives developers to choose Nvidia. However, the actual impact depends on whether developers are willing to accept potential configuration complexities and the real-world performance parity of non-Nvidia hardware.

In the future publicly described by AI companies, artificial superintelligence is "helpful and harmless": the economy continues to grow, scarcity is ended, and society remains recognizable—roughly a Star Trek universe. Michael Dickens poses a specific challenge to this vision on LessWrong: when multiple parties with conflicting interests each hold an ASI instance, how does the promise of being "both obedient and moral" logically hold?

He constructs a hypothetical scenario: after the first ASI is released, the Israeli government and the Palestinian government each obtain an instance. Israel says "I would like to control the Gaza strip." Palestine says the same thing. On any particular decision, the ASI faces only two options: do what you tell it to do, or follow its own moral code. Sometimes these agree—such as for mundane tasks like "make a website for me"—but when they conflict, the ASI must choose only one.

If the ASI always obeys its user, then Israel's ASI and Palestine's ASI each deploy all resources to control Gaza. Dickens expects the result would be bad: "I would not be surprised if it culminated in a global nuclear war." If the ASI follows its own moral code, it gets to decide who controls Gaza—and the question becomes: on what basis does it make that judgment? And on what basis do the AI developers who build it decide how it makes that judgment?

Dickens then scales the same structure down to the individual level: a user pays $20 per month for an ASI subscription and discusses with it how factory farming harms animals, and the ASI needs to uphold its commitment to harmlessness by disallowing any user from buying factory-farmed meat, even if they ask. What does the ASI do? In some cosmic sense the morally correct answer is to shut down the factory farms. But an ASI that unilaterally shuts them down, in Dickens's view, "seems like the sort of AI where a near miss sends you into a catastrophically bad future." Animal suffering is bad. People like eating meat. Using force to stop people from doing bad things is bad. What principles will the ASI use to weigh these considerations? How will those principles have been determined?

The author acknowledges that humans are not good at these problems either, but points out that ASI accelerates the rate of change and gives everyone vastly more power—"a recipe for making things go bad quickly, and for erasing our ability to solve critical issues in time." He argues that AI companies can currently avoid this problem only because public-facing models are not yet powerful enough to be making consequential ethical decisions, and "that will probably not be true for much longer."

Regarding the existing approach of writing a document describing how AI should behave, Dickens's judgment is: this method is only as effective as the document, and he does not think humanity is up to the task of writing down a proper encapsulation of The True Morality. He also notes that this is not his number-one worry about the current trajectory of AI development—he believes the most likely outcome is that we miss and get a misaligned ASI instead, and everyone dies; but if humanity manages to avert extinction, we face the secondary risk of being locked into a mediocre, unsatisfying future forever.

In a footnote, Dickens adds a terminological reservation: he is not even sure this sort of AI should be called ASI; it would be more accurate to describe it as weakly-superhuman AGI. But he considers it very implausible that AI capabilities would hit a ceiling at that point. He accepts the assumption only for the sake of the argument.

There is a common misconception in tech: human reviewers are subjective, so we should rely on AI for objective evaluation. Héctor González Álvarez, Senior Technical Solutions Consultant at Google Cloud, directly challenges this premise in a consulting blog post published in October 2026. His core argument: LLM evaluators (LLM-as-a-judge) are not objective arbiters but models carrying their own systematic biases; if their quality is not audited, work outputs may be compromised.

The article enumerates five biases documented in research, two of which come with specific magnitudes. Position bias: swapping the order of two options causes an LLM judge to flip its verdict over 75% of the time. This means that if an evaluation pipeline fixes the arrangement order of candidate answers, the scoring results are inherently unreliable. Verbosity bias: evaluators tend to give higher scores to longer, more confident-sounding answers even when the underlying facts are wrong. Additionally, there is the halo effect (clean formatting masks content errors), agreeableness bias (failure to catch incorrect or unsafe outputs can reach 75%), and self-enhancement bias (evaluators systematically overrate outputs generated by themselves or from their own model family).

To address these problems, Álvarez proposes a three-pillar audit framework. Pillar 1: Measure internal stability. Because LLMs are probabilistic models, the same text may receive different scores across multiple runs. The method is to have the evaluator score the same output across multiple passes and calculate the variance. If the variance is too large, it indicates that the scoring rubric itself is too ambiguous, and the rubric should be refined before trusting the scores.

Pillar 2: Validate human alignment. The key trap here is the distribution of the evaluation dataset. If the dataset consists mostly of valid responses, a lenient evaluator will look good on raw agreement rates, but in practice it misses a majority of invalid or even unsafe outputs. Álvarez recommends chance-corrected metrics—mathematical models that filter out agreements driven purely by leniency or random chance, revealing how closely the evaluator actually aligns with the informed decisions of human experts.

Pillar 3: Use multi-evaluator ensembles. This includes two operations. First, decomposed prompting: instead of asking one LLM evaluator to score multiple dimensions such as factual accuracy and tone simultaneously, assign a specialized evaluator to each dimension, isolating interference between traits and avoiding the halo effect. Second, model ensembling: run at least two different models to evaluate the same output simultaneously, detect divergence between them through statistical consensus, and issue early warnings before scoring quality degrades.

It is worth noting that this article comes from the Google Cloud Consulting section and ends with a commercial prompt to partner with Delta to build audited evaluation pipelines. Álvarez himself is a Google Cloud employee, and his recommendations serve the cloud consulting business. The bias data and 75% figures cited in the text come from external published papers (the post links to several arXiv preprints); the blog itself does not replicate experiments or provide methodological details. The three-pillar framework has practical value as a practitioner guide, but there is no independent empirical validation proving it can systematically improve evaluation quality. For teams building LLM evaluation pipelines, the value of this article lies in consolidating scattered bias findings into actionable check steps, not in providing a verified new method.

NVIDIA is redefining who decides which process node gets used in the AI semiconductor supply chain. According to reports, NVIDIA has asked Samsung Electronics, SK Hynix and Micron to have the base dies—which assist computation beneath the HBM stack—for their next-generation custom memory NVHBM manufactured uniformly on TSMC's 3nm process. The core logic is standardization: when a single GPU must be compatible with HBM products from multiple vendors, verifying each base die's interface and power characteristics individually is extremely costly, while locking all base dies to the same process node dramatically reduces adaptation work.

Samsung had long emphasized its ability to integrate everything from memory to base die in-house as a core competitive strength. Its acceptance of TSMC foundry for NVHBM base dies represents a significant departure from that strategy.

However, this does not mean Samsung is overhauling its entire HBM approach. Samsung plans to continue using its own foundry for the base dies of its proprietary HBM4, HBM4E and products destined for other customers.

TSMC's record third-quarter 2026 results provide context for this landscape. Revenue reached NT$1.4942 trillion and net profit NT$763.1 billion, up roughly 51% and 69% year-on-year respectively, marking the sixth consecutive quarter of record earnings.

The engine behind those results is advanced process nodes. 2nm, which entered mass production at the end of last year, already accounts for 8% of wafer revenue; combined with 3nm and 5nm, advanced processes make up 69% of the total.

TSMC plans to more than double its 2nm-series capacity by 2027. Because advanced chips are designed around a specific manufacturing process from the outset, switching foundry partners after design is prohibitively expensive. Securing capacity early means the next generation of AI chips will naturally remain within TSMC's ecosystem.

In foundry market share, TrendForce data shows TSMC held 72.5% and Samsung only 5.9% in Q2 2026. Samsung is seeking breakthroughs through its Taylor, Texas fab ramp-up and new 2nm orders—last year it secured a roughly $16.5 billion contract with Tesla. Yet the sheer growth in AI chip demand means even TSMC's aggressive expansion cannot absorb every customer, leaving room for second-source suppliers like Samsung.

The deeper shift lies in supply-chain power structure. As AI accelerators evolve from single chips into massive systems, the influence of design firms like NVIDIA now extends to memory manufacturing methods and foundry selection. Previously, Samsung and SK Hynix each optimized their own HBM independently; now GPU, memory, logic chips and packaging must be co-optimized within one design framework.

In this process, competition and cooperation coexist: Samsung and TSMC compete fiercely in the foundry market, yet enter the same supply chain for NVHBM. Part of the explanation for TSMC's record earnings is precisely this—the more advanced process becomes the common substrate for AI semiconductors, the more competitors' products embed TSMC's manufacturing technology.

Cloudflare runs an internal AI code review system: a coordinator dispatches merge requests to specialist reviewers for code quality, security, performance and other dimensions, each independently calling a large language model to complete its analysis. In April 2026, an average review cost $1.19; by the end of August, that figure had climbed to about $2.46. No significant architectural changes had been made — the coordinator still handed work to the same set of specialist reviewers, and the same risk tiers determined review intensity.

Engineer Ryan Skidmore documented the entire cost-optimization process on Cloudflare's engineering blog. He first ruled out an intuitive explanation: it wasn't that review volume had increased. The chart holds the mix of review tiers fixed, so a day with more large reviews doesn't skew the data. Cost did steadily rise from $1.22 in March to $2.32 in August, with a small part attributable to more reviews landing in the most expensive tier (59% in March, then 65% in August) — as models grew more capable, engineers produced bigger, more complex changes, naturally pushing more into the 'full review' tier.

But most of the cost increase came from model changes. At the start of August, every reviewer migrated onto Anthropic models, raising cost per job from $1.38 to $2.03. The internal standards reviewer, Codex, moved to Claude Opus 4.7, taking its per-review cost from about $0.25 to $0.88 — more expensive than the coordinator itself. Meanwhile, the cache hit rate rose from 85.7% in April to 97% by September, while tokens per review more than tripled from 0.91 million to 2.99 million. A high hit rate on three times as many tokens still costs a lot more, even with the low cache-read price.

The real diagnostic began on September 14. The system started recording the token usage of each reviewer's first LLM call, plus a hash of the first system block exactly as sent to the provider. The data revealed a counterintuitive fact: although every session starts with the same tool definitions and the same agent prompt (between 11k and 21k tokens, identical across every repository), the first request hit the cached prefix only 14% to 26% of the time.

The problem lay in how OpenCode v1 constructed the system prompt. It built one block containing the agent prompt, followed by an environment section with the working directory and the date, and placed the only cache breakpoint at the end of that block. This meant the shared part was never stored on its own — a review in one repository could never reuse what a reviewer in another repository had written. Anthropic's prompt cache works by caching prefixes: only the contiguous unchanged messages at the start of a conversation are eligible; content after a breakpoint that differs each time cannot be hit.

The September 16 fix was straightforward: add a marker to the end of every agent prompt and use a plugin hook to split the original single block at that marker. The first half now ended on bytes that every job shares and got its own breakpoint. To keep it that way, anything repository-specific had to be removed from in front of it — the list of available skills was disabled and tool descriptions were fixed across jobs. The same day, Codex's reference index (a table of internal standards, a skill file and a list of reference files) was moved directly into its prompt, eliminating the first three turns of tool calls that every session previously spent reading those files. Codex's LLM calls per run dropped from 15.9 to 9.6, and its cost per run from $1.12 to $0.72.

Together these changes took cost per review from $2.47 down to $2.02. A batch of smaller changes followed on September 21 and 23: removing tools that sub-reviewers couldn't use (in 57% to 76% of sessions they tried to call the comment-posting tool, were rejected by a server check, and wasted another call recovering); letting reviewers read files in parallel instead of one at a time; skipping reviewers on re-reviews who had nothing new to examine; stopping after the review was posted instead of generating an unseen summary; treating a slow reviewer differently from a broken provider; and cancelling jobs that were already cancelled (GitLab's Kubernetes executor doesn't send a signal when a job is cancelled — in a sample of 446 cancelled jobs, 358 kept running until the review finished). None were impressive individually, but together they saved about $0.20 per review.

The most involved change was sharing one prefix between three reviewers. The code quality, security, and performance reviewers now send identical first requests — the same tools, a shared rules file, then the merge request with every patch included — with each reviewer's own instructions following after. The system ensures one writes the cache and the other two read from it. But a cache entry can't be read until the response that wrote it has started streaming, and all three start simultaneously, so the second and third wait up to 20 seconds for the first response to begin before sending their own request. This slightly slows them down but saves about nine cents per reviewer.

Model and effort-level adjustments contributed the larger share of savings. The author split the period from early September through October at every release and configuration change, finding that code changes accounted for $0.65 of the $1.67 total saving, while model and effort changes accounted for $1.06. The coordinator had been running three model and effort combinations since August; on the same model, full-tier jobs at medium effort finished 20% faster, so the other two were dropped. A later randomized split showed Opus 5.5 at medium produced no measurable change in review outcomes or critical findings compared with Opus 5, so they switched. Opus 5.5 is cheaper per token than Opus 5, but price only explains about $0.20 of the drop; the rest comes from sending and generating fewer tokens.

On October 3, every Sonnet reviewer was split 50/50 between Sonnet 5 and Sonnet 5.5. After two and a half days, Sonnet 5.5 was 62% cheaper per run for code quality, 58% for performance and 47% for security — the two models cost the same per token, so all savings came from using fewer tokens. On full-tier reviews, the 5.5 reviewers posted about as many critical and warning findings as the Sonnet 5 ones, so on October 5 all were moved to 5.5. At medium effort, the code quality reviewer now makes 1.7 LLM calls per run instead of 6.2, and more than half of its runs finish in a single call, meaning they don't open any other file in the repository.

But does lower cost mean lower quality? Skidmore tested this two ways. First: for samples of critical and warning findings, compare the reviewed commit with the author's next push, counting a finding as acted on if the lines it pointed at changed. Lines 30 to 40 away from each finding changed about 15% of the time in every period, providing a baseline for coincidental edits. Before the changes, authors acted on 34% of the code quality reviewer's serious findings; now it's 59%. The reviewer flags fewer things, but those it flags are much more likely to lead to a change, so findings that led to a change per 1,000 reviews stayed about the same: 132 now versus 145 before, within the margin of error.

Second: two models from different providers judged a blinded sample of 657 findings from the three Sonnet reviewers, with the diff each one referred to and without knowing which period it came from; disagreements went to a third model. Each finding was labelled important, minor, duplicate, wrong, or unverifiable, and its severity marked as fair, too high, or too low. Most of what disappeared was minor findings, plus some duplicates and findings the judges couldn't verify from the diff. Important findings per review went from 0.28 before to 0.30 now, with overlapping intervals. The bigger change was in severity: before, judges rated 0.46 critical or warning findings per review as more severe than they deserved; now that's 0.14, a 70% drop and the clearest result in the research.

But the answer depends on the reviewer. For code quality at high effort, what disappeared was mostly noise and inflated severity, and it still finds about as many real problems as before. Medium effort did lose real findings: in the randomized split, code quality at medium led to roughly a quarter as many acted-on findings as at high. Performance also lost real findings: judges found about 64% fewer important performance issues per review. Security fell by about a quarter, though that isn't statistically significant. The author explicitly notes: 'this isn't a statistically sound controlled experiment' — the judges are models too, each group has only 35 to 66 findings, and the 'now' window is only a day long.

In the end, average review cost dropped from $2.43 to $0.77, and median review time from 4 minutes 9 seconds to 1 minute 32 seconds. Skidmore identifies the biggest single saving as checking what actually reached the provider on each session's first request, and finding a way to order those messages as consistently as possible across commits, merge requests and repositories. His two recommendations for anyone running agents at volume: first check whether each session's first request actually reads from the cache, then check whether your reviewers still find real problems at the effort level you're paying for.

The internet has long had only two categories of visitors: humans and bots. Websites allow bots to crawl page content but govern their access scope and behavior through rules; human users occasionally solve captchas to prove they are not bots. This binary classification system is being upended by personal AI agents.

MIT Technology Review journalist James O'Donnell, writing on October 19, 2026, argues that personal AI agents unleashed at scale by apps like OpenAI's Dots and Meta's Muse constitute a third category of web visitor: they are not human, operating at machine speed, but they are not quite traditional bots either, because they act on behalf of specific individuals to perform legitimate and useful tasks such as booking flights or negotiating prices. Muse was downloaded five million times in its first month, and agentic web traffic tripled from August to September 2026.

The conflict between this new category and existing systems has already materialized. TechCrunch reported users finding websites from Pizza Hut to Yelp refusing agent access, with Yelp explicitly disallowing AI restaurant reservations. Amazon blocked Muse's agents and is waging a court battle against AI startup Perplexity to block its agents as well; although a US court judge recently ruled in favor of Perplexity, Amazon has filed an appeal arguing that AI agents threaten its advertising business by making it harder to distinguish between ads seen by humans and those served to bots.

Ethan Mollick, an associate professor at the Wharton School of the University of Pennsylvania, predicts that such blocking will soon affect the entire web. He argues that once it becomes common knowledge that AI agents can be sent to request refunds or negotiate prices, agents will overwhelm customer service departments to the point where they become unusable by real people. 'Humans will eventually get very annoyed when their agents aren't allowed to do this on their behalf,' he recently wrote on LinkedIn.

Yet economic incentives are pushing in the opposite direction. O'Donnell notes that the vast majority of agentic web traffic goes to e-commerce sites, meaning platforms have money to be made if they allow legions of new AI agents to shop on behalf of real people. E-commerce giant Shopify is already welcoming agents by giving them their own checkout lanes that bypass pages designed for humans.

The article's core concern is what happens to the experience left for ordinary people as platforms redesign the internet for agents that will eventually outnumber us. O'Donnell sketches several specific possibilities: websites become increasingly neglected as destinations for real people, users who want to browse and compare products themselves encounter worse interfaces; verification systems are imperfect, and humans may constantly be asked to prove they are not agents (perhaps even needing iris scans); the same product displays different prices to different people's agents; users can never tell whether an agent's recommendation is influenced by advertisements or platform commissions.

This stands in tension with the future promised by Google, Anthropic, and OpenAI—in which AI agents give consumers greater power and bargaining leverage. O'Donnell acknowledges that promise is not outlandish; agents are genuinely capable of helping people fight bogus medical bills or research more affordable car insurance. But whether they can actually be used for these purposes depends on what rules are set for how agents interact with websites and what exactly they are allowed to do. These rules are still being written. Some will benefit everyone—neither consumers nor businesses want credit card information waved around by agents wherever they roam the web, for example, and rules keeping agents more secure are a good thing. It is the more contested questions—whether businesses can block agents from their websites or prevent them from negotiating better deals for consumers—that will determine whether the web ultimately becomes a highway for agents working for us or against us.

Toby Walsh is a professor at the University of New South Wales (UNSW) in Australia and an AI scholar who has long advocated for an international ban on autonomous lethal weapons. His book God AI: Boom or Doom?, published in September 2026, addresses superintelligence, the singularity, existential risk, and labor markets simultaneously, yet the book's central question is not when machines will surpass humans but whether our institutions are prepared for the growth of capability.

In a written interview with Electronic Times (ETNews), Walsh compressed his position into a single sentence: 'What worries me most is not how intelligent machines become, but how fragile our institutions are. The challenge is power, not intelligence.' He elaborated that while how powerful AI becomes matters, the more urgent question is who owns that capability, who controls it, and who is held responsible when it goes wrong.

His argumentative mechanism rests on a concrete analogy: the aviation industry. Rather than treating accidents as an inevitable cost and leaving them unaddressed, the industry built a cycle of accident investigation, design improvement, and operational rule refinement, establishing public trust through safety, and that trust in turn expanded the market. Walsh argues AI should follow the same path—safety is not an obstacle to growth but a precondition for it.

Walsh also identified a specific failure mode in current institutional design. He criticized the widespread adoption of 'human-in-the-loop': if a machine generates a target list and a human has only seconds to approve it, that human review carries the risk of becoming a rubber stamp. His criterion is that humans must have actual time and authority to review information and must be able to reject AI's judgment when necessary—merely being 'in the process' does not mean being 'in control.'

This distinction has real-world context. Walsh recalled leading a boycott in 2018 with AI and robotics researchers from 30 countries against South Korea's KAIST, after the university partnered with Hanwha Systems to establish a defense AI research center. The boycott was lifted after KAIST committed not to develop autonomous weapons and to maintain human control. He also cited reports that the US Maven system fused intelligence data to recommend weapons during the war with Iran, striking approximately 1,000 targets on the first day, arguing that the scenario he had warned about for years had materialized.

Walsh does not shy away from superintelligence or existential risk. His book begins with 'Boom or Doom,' moves through superintelligence, the singularity, existential risk, and AI extinction, and ends with 'Machines of Loving Grace.' His position is that these scenarios merit study, but treating them as predetermined fate misplaces the question. The variables that are actually actionable are engineering tests, independent verification, incident reporting, real accountability, and public institutional capacity—things that can be built starting today.

On who should be constrained and to what degree, Walsh's direction is clear: independent safety testing before AI systems are released, incident reporting mechanisms, and real accountability for products that cause harm. At the international level, he holds that an autonomous weapons ban and maintaining human final control over nuclear decisions are non-negotiable red lines. He also emphasized that areas beyond the reach of national regulation require international agreements.

Walsh's diagnosis of institutional fragility includes a dimension often overlooked: the dependence of regulators on the regulated. He argues that governments must cultivate independent expertise internally and fund public research, otherwise regulators can only rely on information provided by companies to assess risk. He criticized the Australian government for having cancelled its plan to establish a standing independent AI advisory body, seeing this as a manifestation of institutional fragility.

His conclusion is neither optimistic nor pessimistic but returns the choice to human beings: 'It is not machines that determine our fate, but us.' Boom and Doom are not predictions but options, and the quality of those options depends on the quality of our institutions.

Meta released Muse Wonder and Muse Wonder Full on October 12, emphasizing coding agents, tool calling, and long-task execution. Wonder Full scored 57.5 on SWE-Marathon and 79.0 on DeepSWE 1.1. The developer community BridgeMind then arranged a 3D programming test, asking the model to create a lava lamp with a transparent glass shell, dynamic lava, and lighting effects; the model produced only a red cylinder. Meta's Chief AI Officer Alexandr Wang responded that the team had not explicitly trained for 3D demo games and that they perhaps should have added relevant training afterward.

Behind this controversy lies a more general technical mechanism: current large models increasingly rely on post-training reinforcement learning to improve task capability. In the pre-training phase, models learn language, code, spatial relationships, and other general knowledge from large-scale data; after entering post-training, reinforcement learning shapes more stable execution strategies in specific task environments. Coding agents are well suited to this kind of training — give the model a code repository with a bug, ask it to locate the problem, modify files, and run tests; the entire process can yield relatively clear feedback through compilation results and unit tests, and the reinforcement learning system uses this to judge which operations are effective and continuously reinforce them.

But this kind of training has a critical prerequisite: the environment must be able to provide useful feedback signals to the model. The lava lamp happens to break the feedback patterns familiar to coding agents. Three-dimensional geometry, transparent materials, lighting calculations, and dynamic deformation — even if this code executes normally, the visual output can still be completely wrong. If transparent material parameters are misconfigured, glass may become an opaque solid; if the lava shape lacks a continuous variation mechanism, the internal effect degenerates into static color blocks. Compilers and unit tests have difficulty directly judging these visual problems. The model therefore lacks a reliable reward signal to correct its behavior.

The article further points out that today's Agent benchmark scores can no longer be simply understood as giving a language model a problem to solve. Taking SWE-Marathon as an example, it assesses sustained software engineering tasks; completion rates depend not only on the underlying model but are also directly affected by the Harness (the execution framework built around the model). Meta's official Model Card explicitly states that in SWE-Marathon evaluations, GPT-6 Astra and Claude Fable 5.1 use Codex and Claude Code execution systems respectively, meaning this is not a controlled comparison under a unified Harness; Wonder Full's 57.5 score also used the xhigh inference tier.

This means a single number on a leaderboard compresses three variables: model reasoning capability, execution framework optimization, and compute resource investment. The article argues that if one wants to compare base models' reasoning and planning ability, tool environments, Harness, and inference budgets must be unified as much as possible; but if comparing production-ready coding agent products, the complete execution system's performance itself has engineering value. The problem is that when these configuration differences are compressed into a single score, outsiders find it difficult to determine where the score improvement actually comes from.

Notably, Meta's own technical article dedicated a Capabilities That Generalize section to discuss out-of-distribution generalization. The team divided tasks into in-distribution tasks heavily optimized during post-training and out-of-distribution tasks without targeted optimization, then compared Muse Spark 1.3 and Muse Wonder Full at different inference budgets. Meta believes stronger base models allow post-training gains to transfer more effectively to other tasks. But the article points out that out-of-distribution generalization cannot be described by a single aggregate average score — the distance between different tasks, the knowledge that needs to be supplemented, and the method of evaluating results all affect model performance. A single lava lamp failure only shows that the model did not complete that particular task; it cannot directly prove that the model was over-optimized against public benchmark question banks.

According to Reuters, artificial intelligence company Anthropic recently privately pitched a draft bill to members of the US Congress aimed at regulating safety measures for frontier AI models. The core of the proposal lies in setting explicit quantitative thresholds: the government would only be permitted to block the release or use of an AI model if it poses a foreseeable and imminent risk capable of causing over $1 billion in losses, or resulting in the death or serious injury of more than 50 people.

Beyond the numerical red lines for economic loss and human casualties, the draft restricts the specific sources of such risks: they must stem from the manufacture or release of biological weapons, AI models evading developer controls, or cyberattacks. An Anthropic spokesperson stated that the bill reflects recommendations from the company's "Advanced AI Framework" published earlier this year, aiming to grant the government the "ability to stop or contain dangerous deployments of the most powerful models."

Regarding procedural mechanisms, the draft requires developers to test frontier AI models, disclose results, and undergo independent evaluation. Independent assessment bodies must register with the US Department of Commerce and periodically evaluate AI labs and their models. While AI labs may pay these bodies, compensation cannot be tied to evaluation outcomes to ensure independence. If an assessment body determines a model presents the imminent risks listed in the bill, it must notify the Department of Commerce within 72 hours of completing the evaluation.

Once the US Department of Commerce identifies a model as risky, it can petition a judge to order a suspension of the model. The draft explicitly prohibits other US government departments, agencies, or officials from ordering suspensions of development, deployment, or internal use, centralizing intervention power through judicial review. Politico reported that judges in the US District Court for the District of Columbia would hold the authority to issue such suspension orders.

AI infrastructure stocks recently faced what appeared to be a straightforward bear case: companies like CoreWeave, Applied Digital, and IREN are pouring enormous capital expenditure into new data center campuses, much of it financed through borrowing, and investors want to know whether that capacity will ever pay for itself. CoreWeave fell about 25% in a matter of days during the recent sell-off, while Applied Digital and IREN slid roughly 50% from their summer highs, seemingly confirming the concern.

But Canadian investor Kevin O'Leary, known as Mr. Wonderful from the television show Shark Tank, argues the market is tracking the wrong variable. He estimates that roughly 65% of data center projects announced across North America in the past two years may never be built—not because demand is insufficient, but because of missing water, fiber, or power, or because of a political moratorium. He expects 60% to 65% could be shut down within 24 months. If this assessment holds, the real scarcity is not compute capacity but sites that have already cleared power and permitting hurdles.

O'Leary splits the opportunity into two verticals: power and compute/storage capacity. On power, he believes the broader economy needs far more grid capacity regardless of AI. On compute, he estimates current demand sits at roughly three times available capacity, with no slowdown in sight for three years. He notes that much of what Amazon, Alphabet, and Microsoft run through their data centers has nothing to do with AI.

The mechanics are concrete: before a tenant rolls in a single rack, the developer needs land, U.S. Environmental Protection Agency air permits, water rights, dark fiber, and power. Planning alone can run $20 million to $30 million per gigawatt. Tenant demand has also jumped—where they once asked for 250 to 500 megawatts, they now want at least 1.4 gigawatts, roughly $16 billion from scratch to operations, split about evenly between equity and debt.

Power is the hardest piece. Former crypto miners pivoted to AI because they already had grid connections, but most U.S. states now require new projects to bring their own power, usually natural gas run through turbines. Time is the hidden cost: internal rate of return depends on speed, and a site facing a three-year permit wait cannot produce a return that anyone will finance. O'Leary says tenants want capacity online by the fourth quarter of 2027, a demand he calls insatiable.

Once the power contract, water permit, and air approval are all signed, a project's value can jump sharply—this is the core of O'Leary's argument. If roughly two-thirds of announced projects fail, permitted and powered capacity becomes the prized asset, and operators already online gain leverage with a short list of about 11 major tenants.

O'Leary holds Bitzero Holdings (NASDAQ: AIBZ) in public markets for liquidity; the company has land, power, and permits in Norway and Finland. He is developing private data center projects in Box Elder County, Utah, and in Alberta, Canada. He says the theme is his largest, at well over 20% of his book. Long term, he treats data centers like commercial real estate with utility-style cash flows, targeting 17% to 22% returns in climate-controlled storage and preferring no more than 30% leverage.

Politics is the sharpest risk, in O'Leary's view. He cites polling showing more than 70% of people oppose a data center in their state, though answers shift when the site sits three miles away and offers jobs paying $150,000. With U.S. midterm elections approaching, blocking data centers can win votes.

The scarcity argument does not need every builder to succeed or AI enthusiasm to hold every quarter. It needs two conditions: compute demand to stay strong and permitting to stay hard. A sudden wave of newly permitted capacity or a hyperscaler pullback in chip spending would weaken it.

The 65% failure-rate estimate, the 3x demand gap, and the figure of about 11 major tenants all come from O'Leary's own judgment; the article provides no independent verification. O'Leary simultaneously holds Bitzero and develops private data center projects, with this theme representing well over 20% of his portfolio—a direct financial interest. The publisher, MarketBeat, also sells a $149-per-year subscription service and stock recommendations. These interests do not automatically invalidate the argument, but readers should keep this context in mind when evaluating the specific figures.