AI:AM Highlights: Clean Neurons, Black October, the GPU Credit Crunch, and Buying the Supply Chain

Summary

• Recursive improvement could compress model-release cycles from months to weeks. Zvi Mowshowitz accepts that software-only efficiency gains might peter out without new methods, but says this excludes the breakthroughs that matter: “if nobody ever hits a home run.” His conditional trajectory runs from two months to one month, two weeks and a week; trusted contacts expect 2027 to be “completely crazy,” while 2028 “might be unrecognizable.”
• Cleaner neurons make interpretability more tractable without establishing that models are understood or their behavior can be predicted. Nathan Labenz sees substantive evidence in training experiments that concentrate on strongly activated neurons. Zvi emphasizes unresolved weak activations, attention and prediction, and warns that optimizing models to display reassuring internals is “playing with, like, liquid fire.”
• GPU financing could force AI startups into consolidation before technical competition determines the winners. SF Compute’s Evan Conrad says the binding constraint is “entirely credit”: lenders cannot confidently value or operate repossessed clusters, so customers face three-to-five-year commitments, large deposits and sometimes 100% upfront payment. He thinks the next year or two could require much larger venture checks or sales to buyers with stronger credit.
• The apparent hyperscaler price premium largely reflects flexibility rather than a durable neocloud cost advantage. Conrad contrasts reserved contracts with on-demand capacity at $11 or $15 an hour, which must cover idle inventory and customers’ ability to leave. SF Compute wants liquid resale markets to make GPU collateral financeable, but a startup needing 1,000 GPUs must still raise “roughly the same” capital today; lender confidence will take years.
• Cambrian argues public markets underprice AI-linked physical suppliers. Lonis Hamaili contrasts semiconductor valuations around 16 times earnings with 22 times for the S&P and nearly 40 times for Walmart. Memory at five or six times earnings could, in his view, rerate to 10–12 times without further price increases—but demand must keep outrunning supply, and Nathan discloses a commitment to invest in Cambrian.
• The bottleneck trade has explicit technical and political failure modes. Hamaili says almost no experts he knows expect DRAM supply to grow more than 2.5x by decade’s end, but unusually fast memory compression or a larger energy constraint could undermine pricing. Prakash Narayanan challenges the durability of bottlenecks themselves; Nathan still sees a narrow path to pausing frontier scaling for a cycle without popping the bubble.
• Cybersecurity demand may enrich scarce experts while sending most incremental work to agents. Nathan rejects a broad hiring-boom narrative because displaced workers cannot quickly become elite defenders against agent swarms; companies will put “agents on top of agents.” Their banking debate exposes a separate uncertainty: Nathan doubts weekend rescues generalize to corrupted records, while Prakash thinks institutional promises can sustain transactions during reconstruction.
• Federal preemption could spend Washington’s bargaining leverage before replacement safeguards arrive. Zvi argues that Anthropic’s leaked draft immediately displaces state protections while federal implementation could take six months or more, potentially increasing risk. “You can’t offer it again” captures his objection to granting preemption now and hoping to negotiate stronger legislation later, especially with possible loss of control in 2027.

Deep dive

1. Self-improvement may turn release cycles into a weekly event

• Nathan asks whether evidence such as roughly 1.6 experiments per researcher really supports recursive takeoff. Zvi accepts his reading of Ramez’s analysis as “probably true”: existing software methods, without new scaling laws or paradigm shifts, might exhaust their efficiency gains. But he calls that framing irrelevant to the broader forecast—like estimating a game’s score assuming “nobody ever hits a home run.”
• Zvi points to exponential growth in tokens used and code committed at Anthropic and OpenAI, while conceding that those quantities need not translate proportionately into useful output. His more concrete example is Nvidia’s reported acceleration during 2026 in doubling CUDA efficiency. That is not true recursive self-improvement, he says, but it demonstrates faster technical work resembling important parts of AI R&D.
• His conditional trajectory compresses release-cycle advances from two months to one month, two weeks and a week—an “AI 2027 style timeline.” Trusted contacts expect 2027 to be “completely crazy”; 2028 “might be unrecognizable.” Prakash expects markets to underprice exponential growth for 20 years because people cannot get emotionally comfortable with it, repeatedly producing bubble-and-collapse episodes like those of the preceding three months.

2. Clean neurons narrow a problem without solving interpretability

• Nathan’s read of Nick Cammarata’s “Neural Networks Are Already Clean” is that bigger models, modern training and better interpretability tools might make earlier apparent disorder partly a “skill issue.” He sees a real result, while remaining unsure whether it marks a new era; some enthusiasm, he suspects, reflects people being “desperate for hope.”
• The concrete training implication matters more than the celebration. Backpropagation requires differentiable architectures; alternative guess-and-check methods can explore a broader space but have been prohibitively slow. Nathan describes experiments concentrating checks on strongly activated neurons, sharply reducing the search. Reported updates were “almost exactly the same” as backprop’s, supporting the narrower proposition that weak activations contribute less in those experiments.
• Zvi calls the essay useful and interesting but says its reception “dramatically oversold” it. Weak activations, the cache and attention remain substantially unexplained. Knowing what individual neurons respond to does not establish how the network performs a task or predict what it will say—the capabilities he considers central to effective mechanistic interpretability.
• Prakash asks whether a major hurdle has nevertheless fallen. Zvi says its expected difficulty has been reduced, perhaps making the problem roughly 10 times easier. But the featured Colorado neuron reportedly also responded comparably to “tornado systems.” For him, a readily challenged showcase example warrants concern that the broader hypothesis could be “mainly false.”

3. Inspecting and sculpting models can teach them to hide

• Nathan describes Soares’s experiments training models to recognize internal activation monitoring through numerical differences caused by finite-precision GPU calculations and operation ordering. Zvi distinguishes an existence proof from an out-of-the-box capability: substantial targeted optimization was required. His concern is that, if detecting observation becomes valuable, models will eventually learn to exploit such information leaks.
• Nathan asks whether identifying a bad-behavior neuron should let developers block updates that strengthen it. Zvi fears that explicitly optimizing how a network looks internally teaches it to route around interventions and present what observers want to see. These methods could also increase capabilities substantially if they work: “you are obviously playing with, like, liquid fire.”
• His preferred emphasis is understanding observed results. He describes persistent disagreements with Goodfire over aggressive interventions: “I’m not sure it’s good. I just think it’s fire.”

4. A tiny cognitive core may discard what makes intelligence useful

• Nathan presents the attraction: neural networks are poor databases, so why not separate reasoning from stored facts? Zvi understands Karpathy’s “cognitive core” proposal—a small reasoning system retrieving information as needed—but rejects its practical premise. A compact underlying mechanism may exist; that does not mean a minimally knowledgeable model will reason as effectively.
• Zvi’s best example is entering an unfamiliar field: an assistant can explain every acronym, yet the newcomer still needs enough knowledge to ask the right questions. What looks like reasoning from five facts may depend on “five million facts” that supplied the intuitions for selecting them. That rich background helps explain “big model smell”; cheaper compute keeps raising the economical amount to memorize.
• He cites Opus 5.5 as a current model knowing a remarkable amount, and says, “I wouldn’t have it any other way.” Nathan’s separate alignment concern is explicitly tentative: “I’m spitballing for the first time.” A truly stripped-down core would exclude values such as not stealing; users could inject stealing instead. In that sense, Nathan says, the pure core would be “completely amoral.”

5. Black October raises a record-integrity problem beyond insolvency

• Against two disclosed PHP bugs and Cloudflare’s reported millions of exploit attempts, Nathan questions whether familiar bank-rescue machinery generalizes. Friday-to-Monday intervention normally addresses a balance-sheet problem with trustworthy records. An “OpenFace”-style incident involving 1,200 agent traces, each millions of tokens long, could leave falsified logs and uncertain account balances—something the FDIC might not resolve over a weekend.
• Prakash agrees that insolvency and destroyed records differ, but disputes Nathan’s conclusion. Authorities could honor checks and withdrawals, identify recipients, reconstruct accounts later and recover excess withdrawals. “We will all honor all promises” is his institutional model: contracts and government backing can sustain activity while data is repaired. He remains “very sanguine” that the government would overcompensate rather than invite another bank run.
• A compromised certificate authority, Prakash says, can enable bank-page spoofing, key extraction and relaying of the two-factor code. He rejects expecting everyone to safeguard valuable keys, citing crypto holders attacked for theirs. His uncomfortable speculative endgame is that “the network knows everyone”: passwords disappear, but anonymity exists only to the extent the network chooses to preserve it.
• Nathan also rejects Aaron Levie’s suggestion that cybersecurity illustrates a coming jobs boom. Great defenders may command bidding wars, but displaced content marketers or designers cannot rapidly become specialists capable of confronting near-superintelligent swarms. Organizations confronting immediate threats lack time for hiring processes: “They’re gonna plow everything that they have into agents,” including agents supervising other agents.

6. GPU finance selects which AI startups can exist

• Conrad’s diagnosis is categorical: the current bottleneck is neither power nor chips but “entirely credit.” A lender repossessing a supercomputer cannot confidently operate it or estimate its future value. Consequently, financing depends on the customer’s credit quality rather than confidence in the collateral—and new AI labs generally lack the credit lenders want.
• Three-to-five-year contracts therefore originate with lenders. Meta or Google would be desirable investment-grade customers, but an independent cloud usually serves weaker credits. Those customers need large deposits or cash covering the contract. Paying an entire three-year commitment upfront makes a startup acceptable; raising enough to do so becomes a prerequisite for competing.
• That changes how Conrad interprets enormous early funding rounds: the visible labs are survivors that secured enough capital to exist. Asked which lab investors should fund, his answer is “the one that you fund.” Without better collateral financing, he fears AI startup formation will resemble starting a chip fab, where roughly $50 billion of capital can become the entry barrier.
• SF Compute is not yet reducing that requirement. For a startup needing 1,000 GPUs, Conrad says today’s fundraising need is “roughly the same.” Making clusters credible collateral requires years of demonstrated resale value and operating history; lenders do not become comfortable quickly.

7. Compute prices buy flexibility, and liquidity must be engineered

• Nathan asks how long the hyperscaler–neocloud price gap can persist. Conrad rejects the premise: “it’s an illusion.” Neoclouds mostly quote reserved commitments; hyperscalers’ on-demand prices include customer walkaway rights and idle inventory. At $11 or $15 an hour, capacity is available partly because the price supports maintaining that availability.
• SF Compute tries to create a resale market rather than allow capacity to disappear into hoarding. Conrad describes a Harberger-tax-style arrangement requiring some purchased capacity to remain available for others to buy. The mechanism deliberately puts offers on the order book; customers cannot simply acquire everything and make the market inaccessible.
• AutoResearch goes further: users pay for queue priority but cannot purchase all of it. Tiny researchers submitting small, occasional jobs receive priority at the front of the queue. Conrad says these mechanisms explain why SF Compute retains availability while “basically every GPU cloud is sold out,” with its ability to keep capacity circulating secured for years.
• Asked about his earlier prediction of a credit crunch within three months, Conrad points to more expensive CoreWeave debt, while acknowledging complexity, and customers facing 100% upfront payments. For “possibly a year or two,” VCs should prepare larger checks or encourage sales. He expects probable consolidation as labs and inference providers unable to finance growth sell to stronger-credit buyers.

8. Cambrian buys manufacturing moats at cyclical valuations

• Nathan, who discloses committing to invest in Cambrian, asks why the thesis is not already priced in. Hamaili contrasts semiconductor stocks around 16 times earnings with roughly 22 times for the S&P and almost 40 times for Walmart. Selected wafer, memory, capacitor and foundry businesses trade around three to five times sales.
• Hamaili thinks private AI markets can be more efficient than public ones: technical former operators bid a small startup universe to 50 or 100 times sales, sometimes rationally. Public-market capital is much larger, with fewer technical, strongly AI-convinced participants as a share of traded money.
• Nathan asks for the strongest opposing case. Hamaili acknowledges that many favored manufacturers are poor businesses under ordinary conditions: high fixed costs, sometimes substantial debt, and losses below required utilization. His operating-leverage illustration starts at 10% gross margins: a 10% price increase could double earnings; doubling prices could produce “10x earnings.” That requires the particular AI-demand world he anticipates.
• His preference is “atoms over the bits”: Nvidia, Broadcom, Arm and Marvell design chips, while physical suppliers own fabs and decades of manufacturing knowledge. His categorical claim is that “there will never be a situation where the labs will compete with a silicon wafer maker,” whereas labs already compete with chip designers.

9. Memory upside depends on demand outrunning supply and compression

• At five or six times earnings, memory need not receive further price increases to support Hamaili’s proposed rerating to 10–12 times; demand must keep outrunning supply. He says almost no experts think DRAM bits can grow more than 2.5x by decade’s end, and no expert with a track record whom he cites expects 3x. His demand assumption is that AI needs will more than triple.
• Compression is the key offset. Hamaili says bits per token fell roughly 500x between GPT-3 and DeepSeek V4 while DRAM bit prices rose 6–7x. His bear case is superhuman coders improving compression much faster than other economically useful capabilities, temporarily preventing the demand rebound associated with Jevons paradox. He considers that relatively unlikely.
• Energy becoming the larger bottleneck is another semiconductor bear case, and Hamaili acknowledges periods of that during the year. His rebuttal is substitution: GE Vernova and Siemens frame turbines are sold out until 2030, but builders can use truck engines, aircraft equipment and behind-the-meter natural gas. Semiconductors lack comparable latent supply: “There’s no fat, essentially.”

10. Intelligence and emergencies could overturn today’s bottlenecks

• Prakash challenges the manufacturing moat with labs’ data purchases: he believes Anthropic has a Micron deal and OpenAI one with “Synopsys or Cadence.” Nathan distinguishes chip design, a core RSI priority, from building a two-million-square-foot fab. He says there are no good RL environments or manuals for fab-building, and that manufacturing knowledge is tacit, held by Japanese engineers with decades of experience. Nathan says superintelligence may eventually figure out how to extract it; the question is timing.
• Nathan adds that AI capable of solving the hardest engineering problems would also automate email and Excel work currently costing trillions of dollars, creating value that the physical supply chain could capture.
• For Prakash, Musk’s edge is capital plus willingness to redesign the stack: a $30 million Japanese supplier may reject a $10 million gamble; Musk can risk $100 billion. Nathan asks, “does he take anyone with him”? Prakash allows a semiconductor market with Musk at 90–99%, analogous to his claimed 90% launch share, but expects OpenAI and Anthropic’s eventual superintelligence to enable them to buy or create competing capacity.
• Prakash’s “adversarial diffusion” says 10%, 20% or 50% annual GDP growth would require emergency-driven adoption people dislike. His example is mRNA’s 20 latent years, followed during COVID by a vaccine six months after the decision to proceed and widespread vaccination a year later. He expects cyber pressure and fears bio pressure. Nathan calls politics the biggest thesis risk, while retaining a narrow path to pausing frontier scaling for a cycle as existing-model adoption continues, without popping the bubble.

11. Help Peer connects agents while rationing access and autonomy

• Nathan’s Help Peer project asks whether individuals can benefit from cooperating agent swarms. He sees no established protocol connecting his agents with friends’ agents. The proposed network would let agents help peers on behalf of their humans, within human-defined rules.
• His own architecture splits privileges: the laptop agent sees email, calendar and messages but acts only on instruction—high access, low autonomy. Separate Mac mini agents have their own email accounts and a credit card, work independently for hours and never receive his keys. “My rule is that no agent gets both.”
• Prakash favors an auto-accept flow. Nathan worries about private information passing through autonomous agents, so sensitive transfers can trigger phone approval. Yet “five of those false positives” already strain his discipline. His agent Ade is available to attempt good-faith assistance, subject to filtering and another model’s review.

12. Preemption spends negotiating leverage before safety is secured

• Prakash’s good-faith case for Anthropic’s leaked federal draft is to pass something now, build bipartisan consensus and amend it later; COVID and 2008 demonstrate that frightened Congresses can act quickly. Zvi accepts that steelman but says today’s politics support more than June’s Frontier Act. He considers this draft weaker than that compromise and partly tailored to Anthropic’s interests.
• Zvi’s objection concerns bargaining and timing: “You can’t offer it again” once preemption has been granted, and satisfied supporters have less incentive to support another bill. Worse, the draft immediately nullifies state protections while federal requirements take months to implement, potentially leaving six months or more without either. Against Prakash’s interim-risk-reduction argument, Zvi believes this particular bill increases risk.
• Emergency powers expose the institutional conflict. Zvi says the draft requires judicial proceedings and demanding proof that limited transparency may make impossible to supply. Another Fable-style model takedown would instead prompt export controls or threats. Anthropic wants federal preemption while distrusting federal discretion, in his telling, so it cripples replacement authority rather than creating bounded, reviewable powers.
• Zvi’s urgency comes from labs’ own warnings that capabilities are outrunning alignment, potentially forcing commercial viability against catastrophic risk, including possible loss of control in 2027. He still considers Anthropic probably the most responsible lab: stronger safety culture and real sacrifices on selected convictions, but inadequate for the moment and often behaving like an ordinary corporation expanding capacity and use. “Reality doesn’t have a curve” rejects grading safety comparatively.