❯ OpenAI and Anthropic publish claimed ETH refutations, prompting a broad scientific review
Two independent claimsOpenAI and Anthropic have released claimed refutations of the Exponential Time Hypothesis, or ETH, produced by internal AI models. In its October 26 announcement, OpenAI dated the discovery to October 16 and Lean formalization to the following day. Anthropic’s repository also contains a paper and Lean 4 code credited to an internal Claude research model.
A longstanding complexity assumptionETH asserts an exponential worst-case time requirement for 3-SAT, a logical satisfiability problem. Researchers have used it to derive hardness results for many other problems. The proposed algorithms achieve subexponential growth. They have not delivered a practical solver or settled whether P equals NP.
Checking and understanding the resultLean checks logical deductions within a formal statement. Researchers must still examine how the statements, computational models and written arguments correspond. Russell Impagliazzo, Ramamohan Paturi and Ryan Williams have organized an open examination, stressing that their involvement does not certify the claim.
Consequences for existing researchIf independently upheld, the refutations would remove the premise behind many ETH-based hardness results. Whether those problems become easier remains a separate question. Practical consequences for cryptography likewise require algorithms that work against particular systems at realistic scales.
▮ SIGNALAs AI produces proofs faster, independent checking, explanation and the search for usable algorithms take on more of the scientific workload.
❯ SemiAnalysis reports NVIDIA inference above 1,000 tokens per second per user on 700B-plus models
Faster interaction with large modelsOn November 1, SemiAnalysis said a new NVIDIA AgentX submission showed HBM-based systems exceeding 1,000 tokens per second per user on models larger than 700 billion parameters, using software written with agent swarms. The reported figure measures generation speed under the test conditions; full tasks also require context processing and tool calls, according to the original post.
Cost depends on the service targetThe post put Rubin NVL72 at 20.9 times GB300 NVL72’s performance per total ownership cost at 325 tokens per second per user, and 9.11 times at 200. Those ratios belong to different serving targets and configurations. AgentX evaluates multi-turn coding-agent workloads, combining responsiveness with serving capacity.
Sharing more cached contextA follow-up post described Mooncake Store pooling KV cache in Vera system memory across 18 compute trays through NVLink, with over ten times the bandwidth of the previous RDMA approach. KV cache retains intermediate context results to reduce repeated computation.
The system determines the gainLarge-model serving costs depend on compute, memory, data movement and software together. Faster cache sharing can reduce time spent waiting for data. Sustained gains across real workloads could let the same infrastructure serve more agents at lower latency.
▮ SIGNALFor long-context agents, inference competitiveness increasingly depends on how cached context moves across memory and compute nodes.
❯ DeepSeek open-sources a Super-Agent prototype for team coding workflows
One repository as the work boundaryDeepSeek has open-sourced dsh-superagent, a coding-agent research prototype. Its official repository describes a system for one GitHub repository, listening to GitHub and optionally Feishu. An authorized request opens a work item, runs in a separate session and directory, and returns a result, such as a pull request, to the original conversation.
Connecting discussion with executionThe design links team conversations, task routing and code execution, reducing repeated transfers of context into a coding assistant. A web panel highlights matters requiring human attention. Actions require a request or a team-written rule, and messages focus on useful results or decisions people must make.
Research code, with operating requirementsDeepSeek labels the release a research prototype, not an official product or production-ready system. The public version differs from its internal implementation, and some proposed designs remain unimplemented. The MIT-licensed code can be reviewed and adapted, but deployment still requires permissions, a runtime and collaboration integrations.
▮ SIGNALOnce coding agents join a team, authorization, context continuity and notification frequency directly affect how much human attention they save.
❯ OpenAI lowers its funding target to about $20 billion at a $1.25 trillion valuation
A smaller fundraising targetOpenAI has cut its latest fundraising target from about $30 billion to $20 billion, while seeking a $1.25 trillion valuation versus an earlier $1.4 trillion, according to Reuters’ account of the Financial Times report. The October 30 report said some investors reduced commitments during the technology selloff, with a first close expected below $20 billion.
Capital alongside compute accessOpenAI earns revenue from ChatGPT subscriptions, enterprise services and model APIs, while requiring compute for research and inference. The report identified a UAE investor consortium including MGX as an expected major participant. Proposed arrangements could also give OpenAI an option to lease new regional computing capacity; participation and compute options remain reported plans.
Competing for investorsReuters said OpenAI wanted to advance the round before Anthropic’s approaching listing, with both seeking some of the same investors. The revised amount and valuation remain fundraising targets, with no final round size announced. Valuation, future capital needs and compute access all shape investors’ funding decisions.
▮ SIGNALFrontier labs’ fundraising negotiations now link equity pricing, continuing capital requirements and access to compute.
❯ Aramco completes a 49% Humain deal, combining about $2.6 billion in cash with AI assets
PIF retains majority ownershipAramco has acquired 49% of Humain through a wholly owned subsidiary, while Saudi Arabia’s Public Investment Fund retains 51%. GamesBeat reported on November 1 that the transaction combines approximately SAR 9.9 billion, or $2.6 billion, in 2026 cash contributions with AI, compute and digital assets.
An integrated AI supplierFounded in May 2025, Humain spans data centers, high-performance compute and cloud, models, and industry applications. It supplies public- and private-sector AI deployments and develops Arabic-language models including ALLAM. Its plan calls for three gigawatts of AI data-center capacity by 2030, a future construction target.
An industrial shareholder and customerAramco will use Humain infrastructure and services as a shareholder and strategic partner for industrial applications. Combining the partners’ assets, talent and investments gives Humain an industrial use case alongside capital. The approximately $2.6 billion covers the cash component, with additional assets also included.
▮ SIGNALAn energy company contributing capital, technology assets and demand brings AI infrastructure development closer to industrial customers.
❯ Meta and Fourier build a humanoid prototype in three months, with a second version underway
Max is an internal research platformMeta and China’s Fourier built a humanoid prototype called Max in roughly three months and are developing a second version. According to Investing.com’s account of The Information report, Meta said the work supports internal research and experimentation, with no plans to sell the units.
Connecting AI with physical tasksThe reported prototype has a facial-expression display, speakers and five-fingered hands for small objects. Work on Max 2 began in September, with planned improvements to its torso, hands and sensors for environments including stairs. Household assistance requires language understanding to connect continuously with perception, grasping and movement.
Hardware partnerships shorten experimentationWorking with an established robotics supplier gives Meta physical equipment for testing and training sooner. The prototypes explore how AI and mechanical systems work together; commercial products would also require reliability, cost and operating-environment work. Meta has announced no consumer launch date.
▮ SIGNALRobotics partnerships give model developers physical test platforms sooner, bringing competition into manipulation and operation in real environments.
❯ Apple CEO John Ternus puts product experience ahead of adding AI for its own sake
Judging features by the user experienceApple CEO John Ternus told WIRED that Apple will use AI to improve products rather than add it for its own sake. He wants people to judge whether they love the experience, without needing to distinguish AI from conventional software. The interview accompanied Apple’s new touchscreen MacBook Pro presentation.
Designing hardware and software togetherA roughly 25-year Apple veteran, Ternus has been CEO for less than two months. The report described hardware engineering and hardware technologies being combined under Johny Srouji, reporting to Ternus. AI experiences depend on chips, devices and software working together. The interview offered no new model or feature launch schedule.
Continuity in the leadership transitionTernus said executive chairman Tim Cook remains available when asked for help. Asked about AI replacing existing applications and product forms, he emphasized human judgment and the personal nature of product creation. His comments place AI’s value in how specific features fit everyday use.
▮ SIGNALAs AI becomes part of devices, users pay for the overall experience, keeping chip, software and interaction design central to product value.
❯ Exin lets an AI agent guide a mouse brain injection, with treatment outcomes still pending
From candidates to laboratory equipmentExin Therapeutics, a neurological drug-development startup, connected an AI agent to equipment so it could select from 19 gene therapies, choose a brain target and coordinates, and guide an injection into a live mouse with epilepsy. WIRED reported that a veterinary surgeon first opened the skull before the agent controlled the needle and injection parameters.
Humans approve irreversible stepsSafeguards included human approval of the injection plan, an emergency stop and a prohibition on changing underlying equipment-control code. The team initially used Claude Opus 5.5, which planned the procedure but declined to guide the needle. Sonnet 5 then carried out the task within the human-organized experiment.
A coordinate failure tests the setupA coordinate-system fault interrupted the procedure. The agent detected an incorrect position but could not change the code under its permissions. A separate agent helped repair it. The demonstration lasted just over an hour, versus about 45 minutes for a typical injection. Seizure outcomes have not been released, and the work has not undergone peer review.
Closing an experimental feedback loopThe demonstration extends AI’s role from reading and analysis into physical experimentation. Connecting observation with equipment control could link hypotheses, execution and feedback. What has been shown is a supervised injection workflow; treatment efficacy remains to be measured, and the demonstration was slower than conventional work.
▮ SIGNALConnecting agents to equipment creates new experimental feedback, while reliability, permissions and human safeguards directly shape the cost of running it.
❯ Mythos 5.5 appears in reported Anthropic site-code traces while the public page still lists 5.1
An early version-name traceTestingCatalog reported on November 1 that Anthropic’s Mythos site code contained the name Claude Mythos 5.5 alongside its cybersecurity and biology-research description. Anthropic has not published a corresponding launch announcement, evaluations, pricing or availability date.
Verified access remains the current offeringAs of November 2, the official Mythos page still displays Mythos 5.1 for vetted cyberdefenders and life-sciences researchers. Fable 5.1 shares its underlying model, with additional cyber and biology safeguards for broader access.
Capabilities and eligibility develop togetherIf a new version retains that structure, its value will depend on both performance and which researchers and businesses gain access. Product details for 5.5 remain unpublished, with no official schedule for a simultaneous Fable 5.5 release. Existing verification programs remain the application route for eligible organizations.
▮ SIGNALFor highly capable specialist models, eligibility and safeguards help determine usable workflows alongside model performance.
❯ GPT-6.2 Astra appears in Codex usage screenshots, without an OpenAI launch announcement
Screenshots show another model identifierScreenshots circulated by Token Gremlin and kimmonismus on November 1 showed the identifier gpt-6.2-astra in Codex usage data and a response fragment. FourWeekMBA’s account said Token Gremlin described a model picker still displaying GPT-6 Astra while some usage breakdowns showed the new label. The screenshots and number of affected accounts remain independently unconfirmed.
No published product detailsAs of November 2, OpenAI has announced no GPT-6.2 Astra availability, pricing or capability changes, and its official news page has no corresponding launch. Kimmonismus predicted an imminent release without a company-confirmed date. Users have no officially announced selection or upgrade path for the model.
A backend label and a user-facing choiceA model picker presents user options; usage classifications attribute service calls. A discrepancy could involve testing, routing or reporting labels, but its cause remains unknown. Any new release would require official details for developers to assess capabilities, costs and workflow changes.
▮ SIGNALBackend routing and user-facing model names can evolve separately, increasing the need for clear version and billing information.
❯ Binance mobilizes about 100 responders after a $6.3 billion theft and restores some withdrawals
Coordinating the responseBinance assembled a global response after roughly $6.3 billion in cryptoassets was stolen. Bloomberg’s account described a war room growing to about 100 people and coordination with crypto companies and the FBI. The exchange must repair systems, replace assets and manage withdrawal demand simultaneously.
The FBI attributes the attackIn its October 31 notice, the FBI attributed the theft around October 29 to North Korean activity it calls TraderTraitor and urged service providers to block related asset flows. The agency previously blamed the group for the approximately $1.5 billion Bybit theft in 2025.
Replacement tokens and phased withdrawalsReuters reported on November 1 that Binance said it had bought replacement tokens covering about 95% of the stolen assets and restored some withdrawals, including bitcoin and ether. Less-liquid tokens made up the remainder. Binance reported $3.64 billion in net outflows as of 15:30 UTC and said withdrawals were stabilizing after an initial spike.
Several recovery processes run togetherReplacement funding came from its user-protection fund, part of its futures insurance fund and corporate resources. Replacing customer backing, recovering stolen tokens and reopening withdrawals proceed separately. Bearing the loss still leaves the exchange responsible for liquidity, reserve disclosure and explaining its operational recovery.
▮ SIGNALA large custodian’s recovery depends on asset backing and its ability to meet continuing withdrawal and information demands.