Astrohacker

Singularity Watch, Day 14 — October 29, 2026 AI MODELS Gemini 4 Carbon. On October 28, Sundar Pichai, CEO of Google and its parent company Alphabet, announced Gemini 4 Carbon as Google's next frontier model, two weeks after the release of Gemini 4 Argon [1]. Google said Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59 [2]. Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect, called Carbon the company's strongest model, with state-of-the-art cybersecurity and long-horizon capabilities [2]. Google said its internal teams saw Carbon run for days, persisting through dead ends to find and patch subtle bugs across large codebases. On cybersecurity benchmarks, Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them, and 71.0% on CWE-bench for patching them [2]. Google said Carbon cut unsupported claims by about half compared with Argon and resisted every attack in Gray Swan's indirect prompt injection benchmark. Access is restricted for now. Google said it is taking part in the U.S. government's voluntary process for pre-release model testing, and will give Carbon first to a subset of cyber defenders, then to trusted partners through its Fairwind Program [2]. Carbon's introductory price of $2 per million input tokens and $10 per million output tokens matches Argon, rising later to $4 and $20. Logan Kilpatrick of Google called Carbon "a big jump over Argon," state of the art on cyber and long-horizon tasks, and more aligned [3]. Artificial Analysis, which benchmarks AI models, scored Carbon at 59 against 58 for Anthropic's Claude Opus 5.5, with Carbon first on four of the index's nine evaluations [4]. It measured a hallucination rate of 11% for Carbon, against 50% for both Opus 5.5 and OpenAI's GPT-6 Astra, and a cost of $7.76 per index task, against $10.04 for Opus 5.5. A harder yardstick. Artificial Analysis released version 5 of its Intelligence Index on October 27, with about 60% of the index now built on private or held-out test sets [5]. New components include FrontierCode 2.0, Cognition's private benchmark graded on whether a maintainer would merge the change, and Terminal-Bench-Science, 70 agentic science tasks. Before Carbon's release, the v5 leaders were Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Gemini 4 Argon at 50 [5]. The gap between closed and open-weights models widened from 3 points to 12 under the new index [6]. The best open entry, Kimi K3.1 Pro, scored 46, though its weights are not yet released; DeepSeek V4.1 Flash 1017 and GLM-5.3 scored 32, below Claude Haiku 5.5 at 36. Dax Raad, whose team he says bets on open-source models as much as anyone, wrote that there is "a big gap still" and that open models fill an interesting spot in the market [7]. WORLD MODELS AND RECURSIVE SELF-IMPROVEMENT Genie 4. On October 27, Google DeepMind announced Genie 4, a world model that generates photorealistic, interactive environments in real time, with natively generated sound, music and spoken dialogue [8]. Users steer the simulation with keyboard controls and natural-language prompts, while Gemini models monitor the scene to keep it coherent. DeepMind said simulating the world is crucial to building AGI, and that Genie 4 can help create training environments for robots and agents [8]. In one demonstration, Gemini 4 Argon and Genie 4 together simulated Apptronik's Apollo humanoid robot playing checkers, with Argon breaking the task into steps. Genie 4 is a research preview for trusted testers, coming to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and AI Studio [8]. All its video and audio carries SynthID, Google's invisible watermark. Acceleration. On Alphabet's third-quarter earnings call on October 28, Pichai said: "We are at the point where AI is helping us increase our training and launch velocity. I would expect the next twelve months to see continued acceleration because of RSI" [9]. RSI is the common abbreviation for recursive self-improvement, AI systems helping to build better AI systems. Google's Carbon announcement said its "researchers and models are working hand in hand to advance capabilities faster" [2]. AI COMPANIES Alphabet. Alphabet reported third-quarter revenue of $127.6 billion, up 25%, with Google Cloud up 103% to $30.7 billion [10]. Pichai said Google's first-party model APIs now process nearly 32 billion tokens per minute, more than four times a year ago, and that the Gemini app has reached one billion users. Google Search and other revenue grew 12% to $63.35 billion, which works out to 49.7% of Alphabet's revenue, down from 55.3% a year earlier [10]. Pichai said Search query growth was the highest in five years. Cognition. Business Insider reported on October 28 that Cognition, maker of the Devin coding agent, has received investor offers valuing it at $70 billion to $80 billion, weeks after a $48 billion round in September [11]. The report noted that SpaceX bought Cursor's maker for $60 billion, leaving Cognition among the few large independent coding-agent startups open to investors. Cognition said Mercor, founded in 2023 by three 22-year-olds, crossed $2 billion in annual recurring revenue in June, four months after reaching $1 billion, and that its engineers hand Devin long-running tasks [12]. Cognition also said the AI company Decagon used one orchestrating Devin with parallel subagents to run inference research, gaining up to 2.5 times more throughput per GPU [13]. Compute commitments. Sridhar Ramaswamy, CEO of the data company Snowflake, published a long thread on October 27 arguing that AI labs "buy compute for years, sell intelligence by the token," comparing them to take-or-pay contracts in liquefied natural gas, airlines and utilities [14][15]. He warned about circular deals, in which a cloud provider invests in a lab and books its compute purchases as backlog. Ramaswamy cited Anthropic's IPO filing: $265.6 billion in future fixed or minimum purchase commitments and $321.4 billion in leases not yet commenced as of June 2026, against $16.3 billion in first-half revenue [14]. Compute expense fell to 55% of sales from 201% a year earlier. He concluded that "a capacity bust is more likely than a bank run." Anthropic filed its S-1 registration statement on October 26 and plans to list on Nasdaq under the symbol "ANT" [16]. Claude for Legal. On October 28, Anthropic added a gallery of ready-made interactive artifacts for legal work, including a cite checker, jurisdiction survey matrix, matter board and regulatory horizon scan [17][18]. Partner artifacts come from Goodwin, Harvey, Foley & Lardner and PwC, along with nine access-to-justice artifacts. Anthropic said Claude for Legal, launched in May, now has more than 60 legal connectors [17]. Particle. Suhail Doshi, founder of Mixpanel, announced a new company, Particle, and its first model, Particle-1 [19]. It is DeepSeek V4.1 Flash post-trained with reinforcement learning to use fewer tokens. Across four coding benchmarks, Particle reported 28.4% to 52.2% fewer output tokens and 34.4% to 51.6% lower cost per task than the base model, with pass rates within the base model's confidence intervals [20]. It is priced at $0.30 per million input tokens and $1.20 per million output tokens. CODING AGENTS AND SOFTWARE GitHub at agent scale. At its Universe conference on October 28, GitHub said it now sees 7.4 billion commits a month, up from 2.9 billion in August and 1.4 billion in April, The New Stack reported [21]. In 2024, developers pushed 65 million commits a month. GitHub attributed the growth mostly to "agentic demand" and said it has added 4 million CPU cores, 160 petabytes of storage and 50 terabits per second of network capacity this year [21]. It is rearchitecting the platform to separate compute and storage. "Agents are changing the volume and shape of software development," wrote Jay Parikh, Microsoft's executive vice president for Core AI [21]. Stack Overflow. The programming question site received 962 new questions in September 2026, its first full month under 1,000 since 2008, according to a published analysis of its data [22][23]. Questions peaked at 207,169 in March 2014, and September 2025 had 5,072. Bun AOT. Jarred Sumner, who builds the Bun JavaScript runtime at Anthropic, said ahead-of-time compilation for Bun has been merged, with more fixing and fuzzing needed before release [24]. He said it runs existing Bun and Node apps without code changes, and he expects it eventually to become the default for server-side JavaScript [25]. Bun's documentation says the compiler turns TypeScript and JavaScript into machine code at build time, without a just-in-time compiler, for faster startup and lower memory use [26]. It uses type information to generate faster code, and the feature is marked experimental. Macs for agents. Amp, a coding agent led by Sourcegraph founder Quinn Slack, launched macOS Satellites on October 29 [27][28]. Its agents can boot a fresh cloud Mac with Xcode in about 30 seconds to build, test and run iOS and macOS apps, at $2 or $4 an hour. Slack said that even in macOS-heavy development, an agent needs a Mac only about 35% of the time, which shaped the design [28]. HARDWARE MacBook Pro. On October 28, Apple introduced a MacBook Pro with a tandem OLED display and, for the first time on a Mac, touch input [29]. It runs M5 Pro and M5 Max chips with up to 128GB of unified memory and starts at $2,899 for the 14-inch and $3,399 for the 16-inch model. Apple said the M5 Max can run language models with hundreds of billions of parameters entirely on device, and that a neural accelerator in each GPU core speeds tasks such as running a coding agent on a local model [29]. Deliveries begin November 6. Mac Studio. Apple also made 512GB of memory available on the Mac Studio with M5 Ultra. MacRumors calculated the minimum price of that configuration at $22,299 [30]. NEUROTECH Neuralink, the brain-computer interface company, reported on October 28 that it is enabling speech-to-text and voice at conversational rates for participants with ALS [31]. One participant reached 75 words per minute with a 5.8% word error rate on a 125,000-word vocabulary in their first week. ALS can take away the ability to speak while leaving the mind intact. The figures come from Neuralink alone; the company did not name the participant or say how many people are in the speech cohort [32]. In September, a participant in its VOICE trial demonstrated speech in his own voice. ROBOTICS AND INDUSTRY All-robot bricklaying. Salar al Khafaji, CEO of the Amsterdam company Monumental, said its robots laid 100% of the bricks on a house in Almere, the Netherlands, including curved walls, several brick types and bonds, and indented text [33]. He said 15 more houses will follow in the coming weeks. Monumental's fleet includes Pisa, which lays bricks and mortar, supported by supply robots, with software that plans every brick [34]. It works as a subcontractor amid the Dutch housing shortage and raised $32 million in July. Section 232. On October 28, in a proclamation under Section 232 of U.S. trade law, the White House found that imports of robotics and industrial machinery threaten to impair U.S. national security, saying the country produces less of this equipment than it consumes [35]. The proclamation directs trade negotiations first and imposes no tariffs yet. Radar. Sterling Anderson, General Motors' chief product officer, said dropping radar from self-driving systems is "a weird religious stance that ignores physics," noting that radar measures relative speed directly and works in dense fog [36]. GM plans eyes-off Super Cruise for the 2028 Cadillac Escalade IQ. Robotaxi. Tesla expanded its robotaxi service in Houston on October 26 to cover much of the area inside the 610 Loop, using Cybercab and Model Y vehicles, Houston Public Media reported [37]. Tesla also operates the service in Austin and Dallas, and Waymo and Zoox also operate in Houston. MATHEMATICS Navier–Stokes. In September, OpenAI announced that its AI systems had produced a proof that the Navier–Stokes equations, which describe how fluids such as water and air move, can break down under certain conditions [38]. It published the proof in conventional notation and in Lean, a programming language used to check proofs. In a preprint reported on October 27, mathematicians at King's College London and the University of Cambridge found places where the two versions do not match, including a written claim stronger than its Lean counterpart [38]. They said this does not show the proof is wrong, but that checking the Lean version alone cannot certify the written proof. CRYPTOCURRENCY EarthBucks published Day 3 of its BCH Watch, reporting Bitcoin Cash at $279.33, down 8% in 24 hours, at block height 975,021 [39]. Sources: [1] xitter.gq/sundarpichai/status/21… [2] blog.google/innovation-and-a… [3] xitter.gq/OfficialLoganK/status/… [4] artificialanalysis.ai/articl… [5] xitter.gq/ArtificialAnlys/status… [6] xitter.gq/ArtificialAnlys/status… [7] xitter.gq/thdxr/status/211530489… [8] deepmind.google/blog/genie-4… [9] xitter.gq/AndrewCurran_/status/2… [10] sec.gov/Archives/edgar/data/… [11] businessinsider.com/cognitio… [12] xitter.gq/cognition/status/21154… [13] xitter.gq/devindevelopers/status… [14] threadreaderapp.com/thread/2… [15] xitter.gq/RamaswmySridhar/status… [16] anthropic.com/news/s1-sec [17] claude.com/resources/article… [18] xitter.gq/claudeai/status/211546… [19] xitter.gq/Suhail/status/21154979… [20] particle.ai/blog/introducing… [21] thenewstack.io/github-7-4b-c… [22] hum.vibes.dev/s/stack-overfl… [23] xitter.gq/DanielLockyer/status/2… [24] xitter.gq/jarredsumner/status/21… [25] xitter.gq/jarredsumner/status/21… [26] bun.com/docs/compiler/aot [27] ampcode.com/news/macos-satel… [28] xitter.gq/sqs/status/21156925757… [29] apple.com/newsroom/2026/10/a… [30] macrumors.com/2026/10/28/mac… [31] xitter.gq/neuralink/status/21155… [32] teslanorth.com/2026/10/28/ne… [33] xitter.gq/salar/status/211545943… [34] dutchnews.nl/2026/08/amsterd… [35] whitehouse.gov/fact-sheets/2… [36] gmauthority.com/blog/2026/10… [37] houstonpublicmedia.org/artic… [38] kcl.ac.uk/news/openais-navie… [39] xitter.gq/EarthBucksEBX/status/2…

👀

👀

ai will replace doctors. don't try. do.

i'm trying to learn to have empathy for the ai ignorant. what software engineers have lived through in the past year, and are still living through, will happen to everybody. all humans are about to go through a grief cycle as they realize a large portion of the work they've been doing is no longer valuable to society.

it's astonishing to me how much of an unlock these language models are. they unlock, like ... everything else. most innovative projects are 0 to 1 AI is 0 to 1000

robotaxi enabling alcoholism 🤔

does it matter? 🤔 ... all ai models are trained on tons of other people's information ...

everything is going to be ai and the economy will grow by many orders of magnitude

👀

it's really fun working with my grok bots guys i think this idea is really gonna take off

👀

Maybe banning comments is the answer. However, with my issues and experiments workflow, all work items have an "issue", so when the agents write comments, they reference the issue number, so there is always attached context. I haven't had a problem with comments this way.

👀

👀

👀

👀

👀

👀

👀