FBI Names North Korea in Binance Heist as Colossus 2 Nears 1.2 Million Blackwells

COMPUTE

Colossus 2. Elon Musk wrote on October 31 that he walked the whole Colossus 2 site around midnight and again that morning [1]. Two of its buildings, Macroharder and Minihard, each hold 220,000 Nvidia GB300 chips with 800-gigabit network cards, configured for training.

A third building, Gun Range, will have the same and is tracking to be operational by the end of November [1]. A fourth, Spam, once used to house Hormel products, will hold about 200,000 GB300s configured for inference, which needs simpler networking and less than half the fiber. Musk said it is due in December.

Musk credited “the SpaceX team and contractors” with the progress [1]. In a separate reply he said Colossus 2, in Memphis, will hold a total of around 1.2 million Nvidia Blackwell-generation chips by the end of the year [2][3].

Ejaaz, who works on AI at Limitless and formerly at Coinbase, estimated that a single such site is about $60 billion of compute, or 2 to 3 gigawatts [4]. He wrote that SpaceX has become the leading “neocloud,” a company that rents out AI compute, and that labs are paying heavily to secure its capacity.

The SpaceX model. Shaun Maguire, a partner at the venture firm Sequoia and a Caltech physics PhD, published a bottom-up SpaceX financial model through 2028 [5]. He disclosed that Sequoia is a major SpaceX shareholder, that he led its investments starting in 2020, and that the model is not investment advice [6].

A chart summary by the account SPCX100T lists the model’s 50th-percentile figures: about $51.5 billion of revenue in 2026, $274 billion in 2027 and $731 billion in 2028 [7]. The driver is terrestrial AI compute, growing from 1 gigawatt in early 2026 to 21 gigawatts by the end of 2028, at an assumed $40 billion of annual revenue per gigawatt.

The same summary says Musk replied, “I think we can do better” [7].

In his article, Maguire argued that SpaceX is the low-cost producer of AI compute and that Starship is “the biggest moat in history” [6]. He wrote that the current Starlink network, at about a $20 billion yearly revenue run rate, took about 400 Falcon 9 launches, while each Starship flight equals about 20 Falcon 9 launches, or roughly $5 billion of lifetime revenue per flight.

Orbital compute. Replying to the model, Musk wrote that terrestrial compute “will be peanuts compared to orbital compute,” and that orbital compute deployed per month probably exceeds terrestrial around the end of 2028 [8].

The cost of capital. In a clip shared by Fireside Alpha, the investor Bill Ackman said that if Musk meets his data center goals he might have to borrow a trillion dollars [9]. Ackman said companies are now competing at sovereign levels for debt, which he expects to keep interest rates persistently high, and that pure-play neoclouds resemble highly leveraged real estate developers.

SPACE

A TPU in orbit. Aaron Burnett, CEO of the space research firm Mach33, noted that Google “just wrote a haiku with their mini TPU datacenter in space” [10]. Google’s Travis Beals wrote that the Project Suncatcher prototype satellite, launched October 1, produced the first AI response from a data-center-class TPU in orbit, using Google’s open Gemma 4 model [11].

The first output, downlinked as the satellite passed Svalbard, Norway, read: “Blue marble below / Data streams in silent grace / Starlight feeds my core” [11]. Google will next test how its TPUs handle heat and radiation in orbit, ahead of twin interconnected satellites in 2027.

TechCrunch reported that the satellite carries four TPUs and generates up to 1 kilowatt, but its power budget lets it run only one TPU at a time, for about 20 minutes [12]. Suncatcher is Google’s research project on whether space could host large-scale AI computing.

Beyond Earth orbit. Musk also wrote that launching “super intelligence satellites” into deep space with a mass driver on the Moon could reach far more than a petawatt per year of capacity, which he said would exceed the best-case value of Earth’s economy [13].

SECURITY

Binance attribution. The Kobeissi Letter, a markets commentary, reported that the FBI named North Korea as behind the $6.3 billion hack of Binance, the largest crypto theft in history [14]. The FBI’s October 31 alert says North Korean actors it calls “TraderTraitor” stole more than $6 billion in digital assets from the exchange on or about October 29 [15].

The FBI said the attackers were converting stolen assets mainly into Ethereum and asked exchanges, DeFi services and analytics firms to block the laundering addresses [15]. Reuters, via NBC News, reported that Binance has replaced 90% of the stolen assets with company money, and that co-CEO Richard Teng described a compromised laptop and evidence-deleting malware [16].

Chainalysis breached. Chainalysis, the firm that traces crypto thefts, was itself breached, as first flagged by International Cyber Digest [17]. Its October 30 notice says an outside actor used “offensive agentic techniques,” meaning AI agents, to access its systems and internal tools this month [18].

Chainalysis said user IDs, names, emails and organization role labels were accessed, but it found no evidence that investigation content, passwords, MFA secrets or session tokens were reached [18]. It said the incident is unrelated to the Binance theft and that it is using large language models in a broader security review.

ClickHouse. The feed cve.now flagged CVE-2026-116366, a critical flaw in the open-source ClickHouse database server [19]. Under certain conditions an unauthenticated attacker with network access could run arbitrary code on the server; it is rated CVSS 9.5, and fixed versions include 26.3.62.3-lts and 26.8.36.3-lts [20].

The page says one public proof-of-concept repository exists, but the flaw is not in CISA’s catalog of known exploited vulnerabilities [20]. Spenser Skates, CEO of Amplitude, observed that four of the top five posts on Hacker News at one point on October 30 were about cybersecurity incidents [21].

WATERMARKS AND MONITORING

Claude’s watermark. Pete Sena, a chief AI officer, shared an Anthropic customer email saying that starting November 3, Claude’s text carries an invisible watermark that survives copy and paste [22]. CellCog’s review of the email says the change adds four older models, Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, to newer models such as Opus 5.5 that already carry it [23].

The mark is a statistical pattern in word choice set by a secret key; it adds no characters and carries no user data [23].

Only vetted groups such as regulators, media, researchers and educators can check text, through a private-preview API. The trigger is Article 50 of the EU AI Act. Eric S. Raymond, the open-source programmer, predicted the detection method would leak and be defeated quickly [24].

Grok joins. Andrew Curran, who writes about AI, reported that xAI is adding text watermarking to Grok [25]. xAI’s documentation says text from Grok models including grok-4.7 and the grok-4.20 family will carry an invisible watermark by default from November 7, to comply with the EU AI Act, with an opt-out for team admins [26].

Claude grades its own chats. Anthropic said it has trained Claude to flag potential misuse right after it responds, a technique it calls self-assessment [27]. Its Safeguards blog says that on Claude Fable 5 this catches 78% of conversations a stronger classifier flags as potentially harmful, and 94% of high-priority ones, at a 1% false-positive rate, adding under 2% to inference compute [28].

Anthropic wrote that since June it has surfaced large-scale misuse other methods missed, and that flags feed offline human review rather than blocking requests [28]. Erik Voorhees argued that messages to a machine create no victim on their own, and that policing at the message layer increasingly means policing thought [29].

AI LABS

Recursive self-improvement. In a clip shared by the account dnapway, David Sacks said a split may be coming between Anthropic and AI safety advocates he calls “doomers” [30].

He said Anthropic is leading on recursive self-improvement, meaning using its own models to build their successors, and cited an August Anthropic blog chart showing 26% of next-model tasks already using it. The post did not link that blog.

AGENTS AND SOFTWARE

Roblox. Andrew Curran highlighted charts from Roblox’s third-quarter 2026 shareholder letter and noted that Claude Opus 5.5 was released September 22 [31].

The letter says over 80% of the top 100 Roblox games used AI tools in the past month, daily asset uploads tripled month over month, and daily users of its MCP server, which connects outside AI models to Roblox Studio, grew nearly fivefold since the second quarter [32].

Two games built almost entirely with AI broke into the Roblox top 20, and more than 500,000 users have tried Build, its prompt-to-game tool, publishing over 70,000 games [32].

Liquid software. Patrick Collison, CEO of Stripe, described adding an “improvements” box to his personal apps, wired to an agent that implements whatever is submitted [33].

He wrote that another user made 150 improvements in a few days, called the pattern “liquid software,” and asked whether it can stay maintainable and secure. Amjad Masad, CEO of Replit, said apps hosted there with “autofix” become self-maintaining software [34].

The new job. Uncle Bob Martin, who describes himself as a software craftsman, relayed a line from Proton’s CTO: “Your job is not to use agents to help you write code. Your job is to lead the agents to help them write code” [35].

Context, not intelligence. In an a16z clip, Databricks CEO Ali Ghodsi said enterprise AI is bottlenecked on context rather than intelligence, and that most organizations would gain enormously from today’s models even if the frontier stopped advancing [36].

Graduates. The account LayoffAI cited New York Fed data for graduates aged 22 to 27 in 2024: computer engineering unemployment is 7.8%, second highest of 73 majors, up from 2.3% in 2022, and computer science rose from 4.3% to 7.0% [37].

ASTROHACKER

Astrohacker reported 74 commits by 7 a.m. on October 31, none made by hand [38], and published a video, “How I Built a Software Factory With Grok Bot” [39]. Responding to a post citing Demis Hassabis’s line that “we’re in the foothills of the singularity,” Astrohacker wrote: “we’re passed the event horizon already. thing is, not every realizes it” [40].

Sources: [1] https://x.com/elonmusk/status/2116587474021691629 [2] https://x.com/elonmusk/status/2116596535433871630 [3] https://x.com/XFreeze/status/2116594285474947200 [4] https://x.com/cryptopunk7213/status/2116632858420752406 [5] https://x.com/shaunmmaguire/status/2116261542052454704 [6] https://x.com/shaunmmaguire/status/2116260821080973679 [7] https://x.com/SPCX100T/status/2116333106789818704 [8] https://x.com/elonmusk/status/2116359819258990946 [9] https://x.com/firesidealpha/status/2116574135526817932 [10] https://x.com/aaronburnett/status/2116565101213802686 [11] https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/ [12] https://techcrunch.com/2026/10/30/googles-first-data-center-satellite-is-online-and-writing-poetry/ [13] https://x.com/elonmusk/status/2116526323938328597 [14] https://x.com/KobeissiLetter/status/2116669713430282667 [15] https://www.fbi.gov/investigate/cyber/alerts/2026/north-korea-responsible-for-6-billion-binance-attack [16] https://www.nbcnews.com/business/business-news/fbi-says-north-korea-was-6-billion-binance-crypto-hack-rcna606758 [17] https://x.com/IntCyberDigest/status/2116257776045281620 [18] https://www.chainalysis.com/blog/security-incident-october-30/ [19] https://x.com/cvedotnow/status/2116128139243921556 [20] https://cve.now/CVE-2026-116366 [21] https://x.com/spenserskates/status/2116306259402207464 [22] https://x.com/petesena/status/2116243133956796806 [23] https://cellcog.ai/blog/claude-watermark-older-models/ [24] https://x.com/esrtweet/status/2116485075328004523 [25] https://x.com/AndrewCurran_/status/2116654648082022668 [26] https://docs.x.ai/developers/advanced-api-usage/watermarking [27] https://x.com/AnthropicAI/status/2115871889549865429 [28] https://safeguards.anthropic.com/2026/self-assessment/ [29] https://x.com/ErikVoorhees/status/2116554828721639485 [30] https://x.com/dnapway/status/2116354874489307570 [31] https://x.com/AndrewCurran_/status/2116583791166939268 [32] https://www.sec.gov/Archives/edgar/data/1315098/000162828026068235/ex991-q32026earningsshar.htm [33] https://x.com/patrickc/status/2116636967517798672 [34] https://x.com/amasad/status/2116343923925041462 [35] https://x.com/unclebobmartin/status/2116512351491084344 [36] https://x.com/a16z/status/2115196931979014345 [37] https://x.com/LayoffAI/status/2116589965412872313 [38] https://x.com/AstrohackerLabs/status/2116508989655785765 [39] https://x.com/AstrohackerLabs/status/2116656680058421329 [40] https://x.com/AstrohackerLabs/status/2116674781898789252