❯ DeepSeek updates V4.1-Flash-1017, improving project development and shortening task completion time
The website, app and API move togetherDeepSeek announced on October 17 that V4.1 Flash had been updated to V4.1-Flash-1017 across its website, app and API. The company says the release improves benchmark performance while reducing task runtime and output length, helping users finish work sooner and at lower task cost. These are results reported by DeepSeek; improvements will vary by task.
The focus is completing whole projectsThe update uses reinforcement learning based on self-feedback to improve the ability to build projects such as games and 3D scenes from scratch. That requires repeatedly writing code, checking results and making changes. An official demonstration presents historical developments through a three-dimensional scene, emphasizing fewer rounds of user explanation and repair when starting with limited prompts and assets.
The surrounding tools matter tooDeepSeek also optimized the model for its Harness, Tencent CodeBuddy and WorkBuddy, and recommends Harness version 0.2.1-rc.1 or later. A harness is the environment that gives a model access to files, execution and tools. The pricing documentation identifies deepseek-flash as the API route to the updated model. Lower runtime and shorter output are separate from a reduction in the price charged per token.
Developers pay for completed workFor teams building software with AI, shorter answers are valuable when code quality and task completion hold up. Fewer detours and repairs can reduce both calls and waiting time across a project. The release’s practical value therefore depends on performance over a complete project, beyond answer length or an individual benchmark score.
▮ SIGNALCoding models increasingly compete on the cost of the whole task, with fewer repairs bringing them closer to the bill users actually pay.
❯ DeepSeek opens Search and Fetch APIs at RMB 5 per 1,000 successful requests, with existing tool calls billed from November
Search and retrieval have their own chargesDeepSeek’s update announcement opens Search and Fetch interfaces, each priced at RMB 5 per 1,000 requests. Search and Fetch requests made through server-side tools when using the DeepSeek API in DeepSeek Harness, Claude Code and Codex have temporarily been free. Billing for those calls begins at midnight Beijing time on November 1.
One interface finds pages; the other reads themThe API guide explains that Search returns structured results for a query, while Fetch retrieves a specified webpage and returns its body in Markdown. Both use the same authentication as the model API. Developers can connect searching, selecting a result and reading the full page into a workflow for agents or question-answering applications that cite external material.
Empty results do not count as successful callsCharges depend on the number of requests, rather than the number of results or the tokens in a retrieved page. A call is billed only when it succeeds and returns search results or page content. Failures, timeouts and searches with no results are free. Model input and output tokens consumed when reading that content remain separately chargeable under the model’s pricing rules.
Web access becomes part of the task budgetTeams can access models and web content through the same account, while also budgeting for repeated searches and fetches. Successful calls accumulate costs when a task revisits the same topic. Query design and content reuse therefore affect the bill, and shorter model output alone does not determine the cost of an entire web research workflow.
▮ SIGNALAn agent’s total cost increasingly depends on the tools it invokes and how often it invokes them, beyond the model’s token prices.
❯ Samsung plans to use TSMC’s 3nm process for NVHBM base dies as Nvidia seeks a common design across memory suppliers
Rivals share a manufacturing chainReuters, citing the Seoul Economic Daily, reports that Samsung has accepted Nvidia’s requirement to use TSMC’s 3nm process for the base dies of custom NVHBM. The newspaper, citing industry sources, says Nvidia made the same request to SK Hynix and Micron. The supply arrangement remains a media report.
The controller moves into the memory stackHigh-bandwidth memory, or HBM, supplies data rapidly to AI processors. In its August technical introduction, Nvidia explains that NVHBM places a custom memory controller in the base die beneath the stack, freeing area on the main compute chip. The design aims to improve data movement, power use and the allocation of chip area.
This order has a specific scopeThe report says Samsung plans to retain its internal production approach for standard HBM and custom products for other customers. The arrangement concerns the base die rather than the entire memory product. The stacked memory and the logic chip beneath it perform different functions, allowing their manufacturing responsibilities to be allocated separately.
Common implementations can simplify integrationNvidia aims to standardize implementations across memory suppliers, reducing the integration and validation involved in changing sources. That could simplify custom-chip development for AI system builders. Memory companies would compete on their ability to deliver a platform customer’s specified architecture as well as capacity and bandwidth.
▮ SIGNALAI platform companies are extending their influence into memory design, creating manufacturing arrangements that bring direct competitors into the same product chain.
❯ Nvidia’s Kimi K3 optimization delivers 1.73 times the decode speed in a specified GB300 test
A faster result under the same configurationNvidia engineers report in a vLLM technical article that Kimi K3’s generation rate increased from 488 to 842 tokens per second using 16 GB300 GPUs, batch size one and a specified DSpark speculative-decoding setup. That is 1.73 times the baseline. The result measures decoding, excludes input prefill and uses a controlled acceptance-length setting.
Less data movement and fewer small launchesThe optimization uses specialized GPU kernels written in CuTe DSL to combine previously separate operations and overlap computation with communication between chips. Repeated launches and data waits slow token-by-token generation. Reducing that overhead can improve inference efficiency while keeping the model weights unchanged.
Open models require ongoing serving workvLLM’s Kimi K3 support guide describes adaptations for attention, mixture-of-experts computation and speculative decoding. The new optimization targets specified hardware, parallelism and batch ranges while retaining necessary production features. The engineers also emphasize numerical correctness, long contexts and realistic workloads so that benchmark gains preserve output quality.
Generation speed is one part of task timeFaster single-user generation can reduce waiting during repeated coding steps. Real tasks also involve reading long inputs, tool calls and multiple turns. A separate AgentX comparison in the article uses different GPU counts and cannot be combined with the same-hardware multiplier. Deployment benefits depend on concurrency, context and hardware configuration.
▮ SIGNALExisting chips still have substantial room for software gains, while production speedups need explicit hardware, workload and correctness conditions.
❯ Nscale faces pushback on a roughly $30 billion IPO valuation as Microsoft exits a UK data-center project
Some buyers want a lower priceBloomberg, republished by The Business Times, reports that Nscale is seeking an IPO valuation of about $30 billion, while some investors favor $16 billion to $19 billion. The company is still addressing SEC feedback on its filing and has not begun formal marketing. Its final offering price and timing remain unsettled.
Selling compute requires completed facilitiesNscale builds data centers and supplies AI computing capacity, combining chips, electricity and facilities into a service customers can use. It raised $3.36 billion in pre-IPO funding in September, but expansion still depends on further financing and project delivery. A private funding valuation can differ from the price public-market investors are prepared to pay.
A UK project loses its main tenantThe Guardian reports that Microsoft has withdrawn from the Loughton project in Essex following severe delays to power availability. Nscale says it will continue the project and seek alternative power sources, and that its other Microsoft contracts remain intact.
Valuation needs delivery evidenceThe IPO disagreement and tenant withdrawal are two concurrent pressures; the reporting does not establish that one caused the other. Both focus attention on when facilities receive power, when customers start paying and how much more funding construction requires. Announced chip counts and agreements still leave a considerable path to revenue.
▮ SIGNALCompute providers increasingly need power connections, delivery and cash receipts to support valuations, as large construction plans alone provide less reassurance.
❯ CoreWeave bond yields reach about 15% as markets reassess Oracle and SoftBank financing risks
Bonds send an early pricing signalInvestopedia reported on October 16 that CoreWeave bond yields had risen to about 15% that week, alongside a sharp share-price decline. Rising yields reflect falling bond prices and greater compensation demanded for risk. They do not immediately reset the contractual interest rate on all of the company’s outstanding debt.
Construction spending comes before revenueA New York Times report republished by The Business Times describes CoreWeave’s timing exposure. It buys chips and commits to long facility leases before selling compute, while customer contracts may be shorter than its own leases. CoreWeave says its financing is disciplined and disputes negative comparisons of its business model.
Each company's pressure is differentThe report also says some debt associated with Oracle data-center projects trades at roughly 80 to 90 cents on the dollar, while the cost of insuring SoftBank debt has increased. The former involves power and approval delays; the latter relates to financing AI investments. These are distinct instruments and market signals, rather than a common interest rate or evidence that a default has occurred.
Refinancing affects the next facilitiesBondholders face price volatility, while borrowers may face higher costs when raising new funds. Delayed facilities postpone revenue even when construction and financing expenses continue. Markets are distinguishing businesses with stable cash flows from those that need repeated new borrowing to complete expansion.
▮ SIGNALThe financing test for AI construction is whether future customer receipts arrive in time to cover today’s borrowing and long-term commitments.
❯ At least 47 large US data-center projects face obstacles, highlighting the gap between planned and available compute
Two figures measure different thingsAI Weekly, summarizing the Financial Times, says RegLayer counted at least 47 large US data-center projects delayed, withdrawn or blocked by local opposition this year. Bernstein separately estimates that about 35% of the planned US project pipeline is credible. One is a project count; the other assesses a pipeline. They do not establish that 65% has been canceled.
Land alone does not make a working facilityThe reporting identifies constraints including grid connections, transformers and liquid-cooling infrastructure. Local concerns about electricity, water and tax incentives can also affect approvals. Financing, land and chips each satisfy only part of the requirements for operation, leaving announced capacity potentially far from near-term available compute.
Project stages need separate treatmentData Center Census tracks proposed, approved, under-construction, operating, paused and canceled projects, tying capacity figures to documents, hearings and local reporting. It cautions that announced capacity represents an upper bound rather than a forecast. Adding projects at different stages overstates the compute likely to become available soon.
Delivery risk reaches customersThese constraints affect cloud supplier selection and launch schedules. Capacity awaiting power cannot substitute for an operating facility. Customers need alternatives for delays, while developers may have capital tied up for longer. A shrinking project pipeline can therefore reflect difficulty converting demand into facilities on schedule, rather than an equivalent decline in AI usage demand.
▮ SIGNALData-center competition extends through the grid, equipment supply and construction schedule before planned capacity becomes compute that can be sold.
❯ Kioxia loses roughly two-thirds of its market value from its June peak as storage expectations are repriced
A steep retreat from the highNikkei Asia reports that Kioxia’s market value has fallen to about one-third of its peak roughly four months earlier. AI Weekly’s account of the report says it briefly became Japan’s most valuable listed company in June. Investors are reassessing financing costs and competition; the valuation decline does not imply an equal fall in current revenue.
Its business stores dataKioxia primarily supplies NAND flash and solid-state drives, or SSDs. Its product background describes uses in PCs, enterprise servers and cloud data centers for storing and retrieving data. AI systems need storage for training material, models and business information alongside compute chips, linking storage demand to infrastructure expectations.
Supply competition meets customer financingNikkei identifies pressure from rising rates on data-center economics and advances by Chinese memory competitors. Storage spans different products and suppliers: NAND flash, DRAM and high-bandwidth memory perform different roles. A share-price decline by itself establishes neither a fall in orders nor the complete replacement of Kioxia by a rival.
Purchases determine earnings deliveryNikkei also says Kioxia will report April–September results on October 30, with substantial profit growth still expected. Current earnings and longer-term market expectations can move in different directions. Customer expansion decisions, purchasing schedules and the effect of additional supply on prices all influence future profitability.
▮ SIGNALGrowing AI demand does not guarantee continuously rising supplier valuations, as markets reassess customer construction, competition and earnings together.
❯ Anthropic proposes allocating 5% of its shares to a public-benefit fund, with further allocations linked to valuation growth
The initial allocation needs shareholder approvalAnthropic announced on October 18 that it plans to allocate 5% of its shares to a new public-benefit fund, subject to shareholder approval. For each 50% increase in company value relative to a specified baseline, the board may allocate another 2%. Further allocations involve board discretion rather than an automatic transfer whenever valuation rises.
The roughly 20% figure includes earlier pledgesThe company says its founders previously pledged 80% of their holdings, while about half of employees participate in a share-donation matching program. Combining these existing commitments with the new fund pledge brings the total to roughly 20% of company shares committed to public benefit. That figure describes commitments, rather than current fund ownership or cash already distributed.
The fund has a different role from internal researchThe new entity will be a wholly owned limited liability company. Possible activities include grants to nonprofits and direct resource distributions to specified groups. Anthropic identifies labor-market disruption as a potential focus. Work with both public and commercial value, including AI safety and applied biomedical research, would largely remain inside the company.
Governance and spending still need detailBusiness Insider reports on the arrangement and its shareholder-approval requirement. Anthropic says it will release details of the fund’s structure, governance and valuation milestones in the coming months. The share pledge provides potential resources, while its practical uses and the timing of spendable funding depend on implementation.
▮ SIGNALLinking public-benefit resources to company value can expand their scale, while governance and actual allocation determine who ultimately benefits.
❯ Anthropic consults lawyers on negotiation rules as it explores agents that bargain for users
Bargaining requires rules as well as skillAI Weekly, summarizing the Financial Times, reports that Anthropic is consulting law firms about agents negotiating prices, service disputes and compensation on users’ behalf. Questions include when to return to the user and how cautiously to treat another agent. The discussions are exploratory rather than a product launch.
An authorization document could establish limitsOne approach described in the report would let users sign a document setting out objectives and constraints. Agents could use it to demonstrate their authority without disclosing all internal instructions. The issue extends beyond bargaining ability to whether an agent can legitimately bind the person it represents.
An internal market revealed capability gapsAnthropic’s earlier Project Deal involved agents buying and selling real items for 69 employees. They completed 186 transactions worth more than $4,000. Participants represented by stronger models obtained better objective outcomes, while users of weaker models did not clearly perceive their disadvantage. The small internal experiment cannot represent an open market, but highlights an information gap.
Less back-and-forth can reduce oversightWhen both parties delegate negotiation, users may save time while accepting worse terms without knowing the relative ability of the models involved. Authorization, reporting and fairness need to be designed together. Reaching agreement is one step; users also need to understand and correct their agents’ decisions.
▮ SIGNALAgent-mediated commerce needs to make both authorization and differences in negotiating ability visible to the people who bear the consequences.
❯ Four Gemini 4 Argon agents complete Minecraft through autonomous collaboration in a 27-hour run
A shared task lasts more than a dayGoogle DeepMind researchers report in a paper released on October 15 that four Gemini 4 Argon agents defeated the Ender Dragon in a randomly generated survival world on Easy difficulty after 27 hours and 14 minutes. They ran 2,783 programs and died 605 times, without human-assigned roles or mid-run guidance.
Code controls the charactersThe agents write JavaScript using the Mineflayer open-source library to move, mine, craft and read environmental information, rather than operating a keyboard and mouse from screen images. The library supplies action interfaces without choosing the next task. Through game chat and shared memory, the agents gradually divide responsibilities such as gathering, crafting, exploration and combat.
Coordination survives repeated failureThe team exchanged 229 messages and recorded coordinates, tool problems and tactical lessons. The paper says an independent observer agent and human review of sampled logs found no cheating. Six server restarts occurred during the run. The result demonstrates sustained coordination and recovery rather than uninterrupted technical operation.
One success leaves reliability unmeasuredThe authors explicitly describe a single run on the first attempt, which cannot establish a success rate. A programmatic game interface also differs from human operation and real-world robotics. The research provides a recorded long task for studying collaboration, memory and persistent execution, informing how more complex software work might be evaluated.
▮ SIGNALLong tasks expose whether agents can retain goals, exchange useful information and resume collaboration after failure.
❯ OpenAI plans humanoid robots, with Altman expecting a demonstration in 2027
A demonstration target comes before public useEngadget, drawing on interviews with Sam Altman, reports that OpenAI will develop humanoid robots. Altman expects an appealing demonstration in 2027, while everyday public use will take longer. The company has not announced the prototype’s form, demonstration details or commercialization date; the schedule remains a management expectation.
A human-like form fits existing spacesAltman argues that doors, tools and workplaces are built around the human body, making a similar form useful for existing infrastructure. Usefulness also depends on perception, instruction-following, action planning and recovery from failure. Explaining how to pick up a cup and reliably performing that action require different capabilities.
Earlier research provides a simulation foundationOpenAI’s Dactyl research trained a robotic hand to manipulate a Rubik’s Cube, transferring control learned in simulation to physical hardware. This was a specific hand-manipulation task, with a limited scope compared with a general humanoid. It nevertheless provides earlier experience in vision, control and differences between simulated and real environments.
Software and hardware must work togetherEngadget’s review of hiring materials identifies work on motion control, actuators, manufacturing and supply chains, with helping skilled workers build infrastructure among near-term goals. A demonstration can show direction; deployment also requires sustained operation, maintenance and reliable task completion. Pricing and service models for household robots remain undisclosed.
▮ SIGNALRobotics requires model progress to translate into reliable physical action and sustained operation in real environments.
❯ Alexandr Wang gains influence in Meta’s AI reorganization as succession discussion remains unofficial
A profile prompts succession speculationExame’s account of a Financial Times profile says Alexandr Wang has received strong support from Zuckerberg during Meta’s AI reorganization, with some people considering him a potential successor. Meta has announced no CEO transition, and Zuckerberg remains in charge. Outside assessments do not constitute an appointment.
From training data to running a model organizationWang co-founded Scale AI, which prepares and organizes data for model training. He joined Meta in 2025 following Meta’s large investment in Scale. The profile emphasizes organization, execution and his ability to communicate with engineers and recruit talent. Those strengths ultimately need to translate into teams that consistently release useful products.
Recent models offer concrete evidenceMeta released Muse Wonder and Wonder Full on October 12, focusing on personal and professional agent tasks. Wonder is available through the Meta Model API and Muse Code, while Wonder Full serves Muse Code subscribers. Meta also plans to integrate the models into its Muse personal assistant, giving observers products against which to assess the team.
Organization reaches users through deliveryMeta says the models can retain important information across sessions, adjust methods during complex tasks and recover from errors. Users experience leadership through dependable memory, execution and delivery. Wang’s responsibilities therefore extend from attracting researchers to repeatedly bringing model capabilities into usable services.
▮ SIGNALAI leadership in large companies increasingly combines talent management, research and product delivery, with sustained releases providing a firmer basis for authority than succession speculation.
❯ Holyvolt raises $45 million in equity to combine AI and physical laboratories for battery-material discovery
New funding supports materials researchFortune reported on October 15 that Swedish battery-materials company Holyvolt raised $45 million in equity, with investors including Zone II and CC.VC. A source puts its valuation at $245 million, a separate figure from the capital raised. Fortune corrected its investor list: Fifth Wall was an earlier investor and did not participate in this round.
Experiments generate the model's evidenceHolyvolt connects AI predictions with physical experiments, using test results to guide subsequent material choices. Wildcat, acquired this year, has developed high-throughput platforms that test battery materials in parallel. Holyvolt’s company announcements confirm the acquisition and research integration. Battery customers would benefit from fewer unproductive experiments and faster identification of formulations worth further testing.
Promising chemistry still needs scaleThe company is developing nickel- and cobalt-free DRX cathode materials, targeting commercialization in 2027. Fortune describes potential licensing, contract manufacturing and joint ventures to scale results, avoiding the need to build every large factory itself. Laboratory performance, cost and consistent manufacturing at scale still require validation with customers and partners.
The advantage depends on repeated testingBattery-material discovery requires physical evidence of cycle life, energy density and manufacturability. Holyvolt’s approach returns model suggestions to the laboratory. Commercial value depends on predictions and experiments improving one another and on the eventual materials meeting customer requirements. The financing gives the company more capacity to pursue that work.
▮ SIGNALAI materials discovery can derive its advantage from accumulated experimental data and validation, with customers ultimately paying for manufacturable materials.
❯ Disney hires more than 60 Character.AI employees, with about $180 million budgeted for three years of compensation and retention
People and technology have separate pricesPPC Land’s analysis of bankruptcy filings says Disney has hired about 65 Character.AI employees, with roughly $180 million in three-year compensation and retention spending. The documents separately list a $30 million retention pool while describing the three-year total as compensation and retention. Adding it again to produce $210 million would risk double-counting.
The estate receives a different paymentCharacter.AI’s bankruptcy announcement says Disney proposes paying $20 million for a perpetual, non-exclusive license to specified intellectual property and related assets, with transition services. Employee compensation goes to employees rather than the bankruptcy estate. The arrangement does not constitute a purchase of the entire company.
The chat platform still needs a path forwardCharacter.AI provides character chat and interactive entertainment, and has entered Chapter 11 while planning to operate during its sale process. Disney’s license excludes the platform, brand and user chat data. The announcement says the platform will wind down if the process fails to find a way to continue it. The transaction remains subject to bidding and court procedures.
Technology value can move separately from serviceThe arrangement separates the value of people and technology from the existing platform’s survival. Disney can absorb development capability while users and creditors await the sale outcome. For businesses purchasing AI services, continuity depends on how teams, data and service obligations transfer, beyond whether the technology attracts a buyer.
▮ SIGNALTalent and technology licenses can carry value into another company while leaving user continuity and creditor recoveries unresolved.
❯ Jack Clark describes accelerating AI progress as Anthropic’s extreme scenario puts 2030 US GDP at $44.4 trillion
The growth argument has explicit conditionsReuters, republished by CNA, reports that Anthropic co-founder Jack Clark argued in an October 17 lecture that AI could sharply accelerate scientific and economic progress. He used the singularity as a framework and cited internal economic modeling. The argument concerns potential outcomes rather than observed economic results or a scholarly consensus.
The $44.4 trillion figure is an extreme pathAnthropic’s economic scenario explorer distinguishes modest, substantial and extreme effects. The extreme path assumes AI outperforms people in almost all knowledge work, acts autonomously at scale and is adopted rapidly, potentially requiring recursive self-improvement. Startup Fortune’s account of the scenarios gives 2030 US GDP of about $44.4 trillion for that case. The official tool uses 2025 prices, limiting comparisons with future nominal GDP.
Capability still needs business adoptionThe model treats economic activity as tasks that can be automated, augmented or newly created. Higher capability still leaves time needed for companies to change processes and workers to learn or move jobs. Anthropic says the tool simplifies reality and omits policy responses, business cycles and some financial shocks. Its results change with the assumptions entered.
Aggregate growth leaves distribution unresolvedUnder the extreme scenario, knowledge-work employment and wages could face pressure while more gains accrue to owners of capital. Clark also called for serious rules. The scenarios help businesses and workers examine adjustment under different rates of change, rather than determining that a profession must disappear or promising an investment return.
▮ SIGNALAI’s potential economic output and the distribution of its gains are separate questions, making explicit assumptions the most useful part of scenario analysis.