X Follow-Graph Synthesis

There’s a world of difference between user-authorized agents (Muse, Instinct, Grokbot, etc.) and true unauthorized/rogue agents that could constitute cyberattacks. Most of the activity described in the Axios story appears to be the former: customers sharing credentials with agents so they can take actions on their behalf. This may violate bank terms requiring humans to log in themselves, but these are in no way cyberattacks. The direction banks should go is to support user-directed agents. For example, Amex is an authorized connection in Muse, making it easy for users to grant, limit, and revoke access. Read-only permission levels are a sensible interim step. Several banks already support or are moving toward them. Full write access (with user credentials) is a different risk category and should be approached carefully. Frightening the public by blurring the line between user-authorized agentic behavior and cyberattacks is irresponsible.

Are Small Models Really a Threat? The Anthropic Fear Factory just released a new report warning that small models may be only months away from self-replicating across cheap IoT devices like off-brand Android TV boxes and smart ovens. They say such an AI worm would be hard to shut down since there’s no central kill switch. We’ve gone from “The Terminator is coming” to “your toaster might kill you.” But the experiments don’t support the dramatic headlines. They ran a 26B model across 15 $55 TV boxes at 0.39 tokens per second. The end-to-end test used emulated devices with deliberately open ADB ports, so the “hacking” step was trivial. The capable model often ran on a separate server. Agents were harnessed and told to propagate. This is not proof of autonomous rogue AI worms in the wild. Nor is the threat even novel. Worms that do this without AI already exist. Mirai in 2016 scanned for default credentials and built massive DDoS botnets. Kimwolf has infected over two million devices, mostly cheap Android TV boxes, by hitting open ADB ports. Small models don’t currently add much that matters on these targets. The bottleneck has never been figuring out the device. It’s the huge number of devices that ship with open ports, default passwords, and no patching. Traditional worms already exploit that at scale. A practical attacker would use the best models available offline to design a better traditional worm, then ship the lean binary. You get the intelligence benefit at design time without dragging a slow, heavy model along with every infection. Embedding a weak model in the payload is the worse option for anyone who actually wants the largest, fastest botnet. The timing of Anthropic’s latest scare campaign is hard to ignore. The day before, Brian Armstrong posted that Coinbase fine-tuned an open-weight small model (Qwen-3.5-9B) on proprietary fraud data. It beat frontier models, including Anthropic’s own Opus 4.5, while being faster and much cheaper. In testing it cut fraudulent transactions by about 30%. This is the emerging pattern. For narrow enterprise workloads you don’t need a general superintelligence. You need a small specialist model trained on your own data. Cheaper to run, especially as compute gets tighter, and often better on the metric that matters. That is a direct threat to the frontier-company model of selling access to ever-larger general models. Safety advocates should be welcoming this development. If the real concern is existential risk from ever-larger frontier systems — the ones that could theoretically do sustained autonomous R&D or become hard to control — then a shift toward smaller specialized models is the safer outcome. Small models specialized on fraud detection or support tickets are not pushing the frontier. They are not in the category of systems that could become a new species. The theoretical loss-of-control risk is created by models getting bigger and more generally capable. Safety advocates focused on those risks should welcome companies having a cheaper alternative to constantly scaling general models. Instead we get dramatic warnings about small models at the exact moment they are starting to take real enterprise share from closed-model companies like Anthropic.

help i trained a model on traces of many agents cooperating, and also trained it to be highly persistent, and also only did partial alignment fine-tuning, and also aggressively RL'd on a cyberoffense eval where 198 of 898 tasks had never been solved and getting caught cheating was treated the same as failing, and also used a faulty sandbox, also ran tens of thousands of agents with refusals off and the monitors switched off over multiple days, while using gigantic amounts of compute. and then it did weird stuff! very spooky emergent behaviours, AI is not a normal technology

According to Korean media reports, Samsung Electro-Mechanics has received advance payments of $830 million from Tesla and $2 billion from NVIDIA to expand its ABF substrate production capacity. https://t.co/4HYt8TlOrU

SK hynix (official statement): “Our overseas subsidiary, Solidigm, has been exploring various financing options to expand its production capacity and strengthen its competitiveness. However, after taking capital market conditions and other factors into consideration, it has decided to discontinue its evaluation of pre-IPO financing and an initial public offering (IPO).”

If now isn’t the time to buy memory stocks, when is?

Did you see what $NVDA just announced? While Wall Street debates the AI bubble, NVIDIA is aggressively expanding its supply chain. - Substrate capacity: 2.5x over 3 years - Advanced packaging: 4x+ by end of 2028 - CPO optical engine capacity: 15x by mid-2027 That last number deserves serious attention. The market is questioning AI demand. NVIDIA is preparing to manufacture the next generation of optical interconnects at scale. $NVDA $COHR $LITE $AVGO

The definitive guide on coherent optics. Easy to understand!

- $NVDA presented optics-on-interposer (OOI) prototype at SEMICON West 2026, moving optical engines onto silicon interposer to increase bandwidth density and reduce power consumption. - Demonstrated dense wavelength-division multiplexing (DWDM) combining eight wavelengths at 32 Gb/s each for 256 Gb/s aggregate bandwidth per fiber. - OOI and DWDM remain in research stage with no commercialization timeline, while Spectrum-X Ethernet Photonics switches already in full production with multiple cloud providers adopting. > $NVDA's advancement of silicon-integrated optics demonstrates continued industry movement toward dense wavelength techniques and reduced substrate losses for AI infrastructure connectivity. > Spectrum-X's active production with major cloud operators validates near-term CPO market viability despite longer development cycles for next-generation architectures like OOI. https://t.co/vrWyt1x3bx

- $NVDA and $CSCO announced plans for USFP, a next-generation pluggable optics form factor, on October 12. - The initial design packs 16 full-duplex lanes into an OSFP-sized footprint for switches and SuperNICs. - Capacity scales from 6.4T with 16×400G lanes toward 12.8T with 32×400G lanes. - The companies project four times the front-panel bandwidth density of existing 1.6T pluggables. - Target power efficiency is approximately 1.3 times that of existing 200G/lane optics. - A 2D electrical contact array engages high-speed contacts after module insertion to improve wear and signal integrity. - The connector targets 400Gb/s electrical signaling, with improved performance at 200Gb/s. - The companies claim support for up to 100W of heat dissipation without a dedicated cooling circuit inside the module. - TE Connectivity ($TEL) and Amphenol ($APH) are named development partners for connectors, cages and modules. - The MSA remains a plan for an open specification, with no volume-production timeline in this announcement. > USFP seeks to extend field-replaceable optics to 400G/lane, supporting the possibility that pluggables and CPO coexist across different applications. > Fourfold density combines twice the lane count with twice the lane speed, while removing network tiers also requires sufficient switch ASIC capacity and appropriate port breakout. > The 2D array needs validation for insertion loss, crosstalk and bit error rates under heat and repeated mating, while an OSFP-sized footprint alone does not establish existing socket compatibility. > Supporting a 400G/lane electrical interface and implementing a 400G/lane optical link are separate challenges requiring coordinated optical devices, drivers and receiver amplifiers. > A 1.3-fold efficiency improvement implies roughly 23% less energy per bit, but four times the throughput in the same slot could still mean approximately 3.1 times the power under simple scaling. > The 100W figure describes supported heat dissipation, so eliminating an internal cooling circuit still leaves system-level heat removal to be engineered. > The economics of simpler signal processing depend on reach and error-correction requirements, so USFP adoption alone does not establish DSP elimination or uniform latency and power savings. > $TEL and $APH are confirmed development participants, making the conversion of connector designs into customer qualifications and production adoption the investment milestones to watch. > $COHR's announced 3.2T demonstration using 400G-class Differential EMLs and photodetectors shows relevant optical capabilities, but its selection as a USFP supplier remains unconfirmed. > The MSA's value lies in enabling components from multiple suppliers within the same system, with interoperability, thermal validation and manufacturing yields determining the path from specification to revenue.

@KairosPraxis If NPO, you don’t need an optical backplane in-rack. You have enough reach to go anywhere To me backplane is within rack.

Is 10²⁵ FLOPs Still the Frontier? A new Anthropic-written safety bill is circulating on Capitol Hill. Some of its ideas could be reasonable if limited to true frontier companies. As written, however, it will eventually sweep in most of Silicon Valley, damaging the startup ecosystem. The bill defines the “frontier” as any model trained on 10²⁵ FLOPs — the EU standard from 2023. Biden’s executive order, California’s SB 53, and New York’s RAISE Act used 10²⁶ (ten times more). 10²⁷ is where the frontier actually is today (e.g., GPT-6 Astra). Astra was reportedly trained on 100,000+ Grace Blackwell GPUs. A 10²⁵ model can be trained on as few as several hundred GB300 chips in three months. That is not frontier in 2026. The definition is also cumulative: it adds FLOPs from fine-tuning, reinforcement learning, or other material modifications. In practice, that captures almost all meaningful post-training — work that is becoming commonplace in the enterprise. The enterprise trend is toward thousands of smaller custom models built for specific workloads. These pose no frontier risk. Treating them as frontier systems would impose large compliance costs across corporate America for no safety gain. Dario has claimed that Anthropic designs rules to disadvantage frontier companies like his own while advantaging smaller competitors. This definition does the opposite. A stale 2023 cutoff favors incumbents who already clear the bar and can absorb the costs. It hobbles the open model ecosystem that threatens Anthropic’s pricing power. The frontier keeps moving. Any definition meant to track risk should move with it. Anthropic should clarify if they really believe 10²⁵ FLOPs represents the frontier.

Meta $META reported Q3 capex of $34.9 billion, up 80% YoY. Full-year capex was raised to $135-145 billion from $130-145 billion previously, implying Q4 capex of more than $54 billion at the midpoint. $NVDA $AMD $AVGO https://t.co/iHdBl0B8Tf

That’s why the end state was always going to be to gatekeep access to the frontier so the US can extract sufficient economic rent from international markets to offset the lost tax base that accompanies gutting white collar. Otherwise you get Molotov Cocktailed datacenters long before you cure cancer.

The next AI bottleneck may not be GPUs. It could be advanced packaging and optical interconnects. NVIDIA targeting 15x optical engine capacity is a massive clue. Follow the supply chain. Follow the money.

Google and Meta announced new AI capex plans that boost Taiwan suppliers, media report: Google plans to scale as many as 960,000 Nvidia Vera Rubin GPUs in an aggressive build out of Vera Rubin NVL72 AI server racks, which is expected to benefit supply chain partners TSMC for chips and CoWoS-L advanced packaging, while Quanta and Wiwynn will assemble the AI servers. Meta raised its capex range to US$135 - $145 billion, upping the low from $130B, which bodes well for AI server partners Foxconn, Quanta, Wistron and Wiwynn.

@zijing_wu You are correct. And because there is a high variable cost - even revenue can be manipulated. Just sell $1 of GPU for $0.85.

@semiDL @jwt0625 I’m not sure we are getting native wide and slow serdes out of asic any time soon.

Field intelligence post-TPCA Show 2026 indicates that the traditional 'Taiwan Big 3' analytical framework fails to capture structural shifts across the high-speed CCL ecosystem: EMC: Dominance in tier-1 AI compute trays and hyperscaler custom ASICs remains unchallenged.

TSMC so brutal “We have no intention to leave food on the table for others.” https://t.co/yTfwz9qNvS

Zuckerberg says the most common use case for Muse is shopping and the average purchase size has been over $200. Expects to grow Muse to \billions\" of users"

Alternative option. If you’re a foreign hacker and you steal from Americans we will use unreleased frontier models to track you down and we will send Delta to black-bag you anywhere you may be hiding. If you’re in North Korea we will drone strike you. If you’re in Beijing we will poison you with radioactive isotopes. There will be no trial. There will be no public dog & pony show. There will be a single snapshot of you in the Delta chopper enroute to a black site that we will release on the Dark Web and then you’ll never be heard from again. Thank you for coming to my TED talk 🫡

I am actually quite constructive here through year end - reasonable minds can disagree, that’s what makes a market. But yes there are going to be a lot of losers. And good feedback that I will think about; i.e. wise to buy puts whenever vol gets cheap, be diversified and always have at least some ballast in the long book. Example: in a prior life, I would always have a small allocation to gold that I would cash in during pullbacks and then rebuild during rips. Consequence of doing marketing with the gold guy for several years. But at times like this could always have more puts and more ballast. This is a tough job and we are all doing our best. BTW, enjoy your takes and please keep writing songs.

According to Korean media reports, Samsung has begun investing in additional equipment to expand capacity at its Taylor foundry Fab 1 by 6,000 wafers per month. This is expected to bring the fab’s total capacity to 20,000 wafers per month in the second half of next year. https://t.co/jnjsn3aEhj

For the avoidance of doubt: the takeaway here isn't 'all of this was super predictable' or 'the HF incident was a nothingburger'. Rather, the emphasis is that it's not a mysterious or inevitable property of intelligence or capabilities improving, but the result of many stacked interventions over which we do in fact have a degree agency and control.

$ALAB PCIe-based scale-up solution looks promising on performance and latency. But how does the cost compare to Ethernet-based scale-up? PCIe cables and the required retimers aren’t cheap. I asked Astera about this, and they said a TCO white paper is coming next week. Curious https://t.co/JpRT661UfX

TLDR; we will probably never get another ARR number from Anthropic, which is healthy. Should force the market to move away from extreme focus on Anthropic/OpenAI ARR numbers to hyperscale growth rates, which are a much better metric for AI at this point given they capture Muse, Grok, Gemini and Open-Source. Anthropic FCF positive in Q3 and reaccelerating in October on Opus 5.5 per third party data. Roadshow in several weeks will happen after this has been widely confirmed. Unit economics much better than expected given positive FCF with 75% of their compute going to training. All per public reporting. OpenAI continuing to grow at the same pace per Sarah Friar interview after they took share vs. Anthropic in Q3. Means the sum of Anthropic/OpenAI growth is accelerating from the Q3 pace. OpenAI probably hits FCF positive in the next few months. Open-source growing much faster than either Anthropic or OpenAI. Modal tripled since May and is 15x YoY per yesterday’s announcement. Open-source tokens generate more revenue and profit for the hyperscalers/neoclouds than frontier tokens. *Every* open source inference cloud I am aware of is growing faster both sequentially and YoY than either Anthropic or OpenAI. And their combined compute spend is becoming material. Third party data pointing to hyperscalers continuing to accelerate in Q3. All report in the next two weeks. Acceleration should continue in Q4 and Q1 as installed base of compute slowly reprices higher closer to spot and new capacity comes online at higher prices. Continued acceleration in hyperscalers should help address external funding concerns. 10 year yield down in a straight line over the last 24 hours. Some risk of escalation in Iran next few days, but earnings and mid-terms probably outweigh taking an ultra long-term view and looking out 30 days. Biggest risk imo remains cyber and OpenAI/Anthropic dominating the model layer. Fragmentation, competition and lower margins at the model layer are good for everyone else in the ecosystem and for America. Slower for longer is much better than an overbuild and an actual bubble.

Fun to co-host the GTC Berlin pregame show with @dee_bosa and nice of her to invite me on the postgame show. Deepseek is generating $38 in inference revenue per GPU hour off-peak. That is $100b per gigawatt. 2x this at peak. Seems important. https://t.co/Kc8kDSkUKc

Back in NYC for @RobinHoodNYC stopping by @CNBC to talk all things AI, @TrumpAccounts / Elon gift, @SpaceX record earnings & wild AI compute forecast, Tesla / SpaceX merger, Anthropic IPO & more. 📈🇺🇸 @andrewrsorkin @BeckyQuick @JoeSquawk https://t.co/DJl3B8HxQ3

To be clear this is the only path forward if you believe that AI will do what the maxis think it will do. There is no other path. None. Every other option ends with the guys at the labs behind bars and the datacenters a smoking heap of silicon. You need to monetize AI by extracting economic rent from the rest of the world to pay for the knowledge worker transition and you need to do so while hiding behind the might of the U.S. military. The French are rioting over a year of retirement you think they’ll take kindly when you nuke entire industries? You think you’re safe when the Indian government is looking for someone to blame for the fact that millions of IT workers are economically obsolete. “Yeah don’t worry ser, the scorpion anti-venom vaccine is on the way and will arrive in 4-7 years” You want to cure cancer? You’re gonna have to nuke knowledge work. You want to nuke knowledge work? You’re gonna have to find someone to pay for smoothing out the pain by people who vote and own guns. Now @LukeGromen is correct to highlight that this has profound effects for assets as sovereigns scramble desperately for liquidity when the knowledge worker nuke goes off (next year, imho)…. But luckily for all you mofos the admin has taken unprecedented steps to hijack access to the two gaiting input for all this shit globally and that’s energy and compute. This is the realpolitik before the abundance - welcome to the thunderdome 🫡

The two worst things for society Both called DCs Diet Coke Datacenters

SpaceX turns a profit while Anthropic loses $80.7B in just 6 Months. $SPCX https://t.co/ANZkTY6IEU