❯ Anthropic offers Blackstone a 10%-plus yield on $60 billion of TPU debt, with Blackstone set to put in $9 billion
Double-digit interest for computeAccording to Startup Fortune, citing an October 16 Financial Times report, Anthropic has offered Blackstone a coupon of more than 10% to invest $9 billion in the junior portion of a $60 billion debt package. Broadcom arranged the package, which pays for Anthropic’s lease of Google TPU chips. The terms are still under discussion and the deal has not closed.
How big the lease isTPUs are Google’s in-house AI chips. The report says Anthropic’s IPO filing discloses a five-year TPU lease worth $125.2 billion. A $42 billion senior loan led by banks covers about a third of it. The junior slice is riskier and paid back later, so it carries a higher rate; similarly rated borrowers pay roughly 7% in public bond markets, the report notes.
Investors run the numbersPrivate credit firms such as Blackstone get returns above public-market levels in exchange for betting that an AI company can keep paying rent. Over the same period, The Information said on X that Anthropic’s revenue growth is slowing and some prospective investors were disappointed by its latest numbers. The slower revenue grows, the more these long leases and high-yield debts weigh on cash flow.
▮ SIGNALFrontier labs’ compute bills are now large enough to be sold in tranches, and paying double-digit interest on the junior slice shows financing costs have become part of the model race.
❯ Anthropic: quantum circuits designed by Claude Mythos 5 cut gate counts by up to 17x, beating the best published designs
Gains on all eight tasksAnthropic said on October 16 that Claude Mythos 5, working from plain-language briefs, designed quantum circuits for eight basic operations across chemistry, optimization, quantum simulation, number theory and numerical computing. Every one improved on published designs, some by an order of magnitude or more. A paper is on arXiv.
Two standout casesFor elliptic-curve point addition, one Claude design uses 12% more qubits but 17x fewer gates. For the decoder in decoded quantum interferometry, its version uses about 12x fewer gates than Google Quantum AI’s earlier construction. Gate count is roughly the number of operations a circuit must run, so fewer gates means cheaper execution on future machines. Claude proposed designs as code while an automated evaluator checked correctness and scored them, with no human input between rounds; the point-addition task ran for 53 hours straight.
Steps still to goThe work came through Anthropic’s STEM Fellows Program. The Crypto Times reports it was led by researcher Andi Gu, with Harvard’s Mikhail Lukin among the collaborators. Anthropic notes the results use a qubits-times-gates proxy rather than real hardware runs, and error correction and hardware control remain unsolved.
▮ SIGNALResearch problems where an automated checker can strictly judge right from wrong are becoming the place where large models most readily produce real results.
❯ Claude traces the origin of NIST’s elliptic-curve seeds in under half an hour, further weakening backdoor suspicions
A two-decade puzzleAnthropic said an internal research model found, in under half an hour, where the seeds for NIST’s five prime curves, P-192, P-224, P-256, P-384 and P-521, came from. The curves are widely used for ECDSA digital signatures, and the unexplained origin of their seeds had long fueled suspicion of a hidden backdoor.
An old random-number functionThe seeds came from the C library’s lrand48 pseudorandom generator, initialized with Unix timestamps between May 1997 and April 1998. The model also explained four binary-curve seeds and the base points of six ANSI X9.62 example curves. The employee who prompted it gave no hint from earlier work.
Clearing, not crackingAnthropic stresses this is not a break of ECDSA and offers no new attack; it makes a planted backdoor less plausible. The B-163 seed, most base points and why some candidate seeds were skipped remain unexplained. The $36,864 bounty for finding the seeds will go to the Internet Security Research Group.
▮ SIGNALLarge models’ first real mark on cryptographic history is helping clear a standard the whole world relies on.
❯ Anthropic publishes internal hiring data: junior share of technical hires falls from 27% to 9% as commits per engineer rise fivefold
A narrower door for newcomersAnthropic published a report on October 16 using internal data to track how AI is changing its own hiring. Since early 2025, junior staff fell from 21% to 14% of technical departments, and juniors fell from 27% to 9% of new technical hires, with junior postings dropping sharply after December 2025. By April 2026, about a third of ML managers and a fifth of software engineering managers said they were hiring less because of AI.
Output up across teamsCode commits per engineer are five times their level six months earlier, and by August 2026 agents worked six hours for every human hour in employee-directed sessions. Sales, legal and finance also handled more work. Over the past year the technical share of headcount fell from 61% to 56%, as sales, finance and recruiting hired faster than engineering.
Limits of the dataThe productivity estimate compares 192 engineers who adopted coding agents with 74 who did not, and the authors warn that adoption was not random, so results may be biased. This is one AI company under the most aggressive usage conditions, and its clearest lesson is the direction: experienced staff get amplified while the entry door narrows.
▮ SIGNALAI is first removing the junior rung of the ladder, and companies will eventually have to say where the next generation of senior staff comes from.
❯ Anthropic pitches its own AI safety bill to US lawmakers, limiting when the government can halt a model
A high barAccording to Reuters via Yahoo, Politico reports that Anthropic privately pitched lawmakers a draft bill on frontier AI safety. Under it, the government could block a model’s release or use only if it poses a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people.
Who assesses, who haltsThe risk must come from a biological weapon, a model evading its developer’s controls, or a cyberattack. Independent evaluators registered with the Commerce Department would periodically assess labs and models; labs could pay them, but not in ways tied to the outcome. Evaluators finding imminent risk would report to Commerce within 72 hours. Commerce could ask a court to halt a model, a federal judge in Washington, D.C. would decide, and other federal agencies would be barred from suspending models on their own.
Drawing clear linesThe draft’s aim is to spell out exactly when the government can step in. Anthropic and the Commerce Department did not respond to Reuters’ requests for comment, and the draft has not been formally introduced in Congress.
▮ SIGNALA safety bill drafted by a company known for safety focuses on limiting government power, showing that what the industry wants most from regulation is predictability.
❯ ChatGPT’s Chrome extension gains computer use, with Work mode filling forms and organizing tabs
Two modesAccording to IT Home via ifeng, OpenAI upgraded ChatGPT’s Chrome extension on October 16 local time. In Chat mode, ChatGPT can read open web pages and answer questions, for example comparing size, price, delivery and return policies across three sofa listings in three tabs.
Work mode takes actionWork mode handles more complex tasks: sending messages, filling out forms and organizing tabs. Examples include filling in a dentist’s form and adding a calendar entry, or analyzing several contractors’ terms and drafting an email. ChatGPT in the sidebar can also pull up the full chat history, so users can continue earlier conversations without leaving the page.
Betting on extensions over browsersThe extension is open to all plans and also works in Chromium browsers such as Edge, Brave and Opera. The report notes OpenAI’s own browser was shut down in August. Users can hand web tasks to AI without switching browsers, a far lower bar than persuading them to adopt a new one.
▮ SIGNALAI assistants are competing for the browser differently now: instead of asking users to move, they move into the browser users already have.
❯ GPT-6 Pro in ChatGPT runs multiple agents in parallel as Codex raises effective usage
One task, several agentsOpenAI announced on X that GPT-6 Astra Pro in ChatGPT got faster responses and better use of plugins and tools. The new feature lets multiple agents work on different parts of a task at once; users can follow each agent’s progress and open finished subtasks to review the work. It is available to Pro users.
Usage and slidesThe same day, OpenAI employee thsottiaux said on X the team raised effective usage in Codex and Work by reusing context across subagents, keeping context cached longer and improving request routing. He also said slides support was added the day before for Pro, Business and Enterprise users.
The cost of parallelismRunning agents in parallel means a task consumes more compute, so whether limits hold up decides whether users dare to use it. OpenAI is pushing parallelism while cutting per-call cost through caching and context reuse; shipping both on the same day is no coincidence.
▮ SIGNALAs agents start to work like teams, the bottleneck shifts from model capability to how much compute each user gets.
❯ DeepSeek upgrades to V4.1-Flash-1017, switching its web, app and API to the new model
Same calls, same priceDeepSeek’s API changelog for October 17 says V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017, with the official web and app updated as well. API usage and pricing are unchanged, and developers can keep calling deepseek-flash to reach the latest model.
Faster and shorterThe changelog says the new version beats the previous V4.1 on benchmarks while finishing tasks faster with shorter outputs. It is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, with Harness 0.2.1-rc.1 or later recommended. The same entry also opens DeepSeek’s in-house search API for web search and page fetching.
Spotted by developers firstBefore the announcement, a developer noticed on X that the deepseek-flash model fingerprint had changed and guessed a new version was live. Shorter outputs matter most to users who pay per token: the same task can cost less and take less time.
▮ SIGNALModel upgrades are increasingly sold on getting the same answer faster and cheaper rather than better answers, which matters more for agents than benchmark scores.
❯ DeepSeek merges four teams into a Product Department led by Harness chief Cui Tianyi
Four teams become oneAccording to The Paper, citing people familiar with the matter, DeepSeek has merged its Harness team with its Web, App and Platform teams into a “Product Department” responsible for all products, design and development, led by Cui Tianyi. DeepSeek has not announced the change.
Where Harness came fromCui studied computer science at Zhejiang University, won six ACM Asia regional gold medals, spent nine years at quant firm Jane Street and joined DeepSeek in March 2026. A harness is the software framework built around a model to help agents carry out complex tasks. His team open-sourced the DeepSeek Harness v0.1 developer preview on August 13, passing 70,000 GitHub stars within 24 hours; the October 3 release added an experimental Claude Code Mods compatibility layer.
From model lab to product companyCui said DeepSeek’s goal is AGI and the Harness team wants to advance it from the harness side, building original products rather than copying existing agent tools. Putting web, app and the developer platform under one leader shows DeepSeek treating the experience around the model as its own line of competition.
▮ SIGNALAs gaps between models narrow, turning a model into a product people enjoy using is becoming the new dividing line between labs.
❯ DeepSeek goes down for nearly five hours and refunds seven days of API spending as credits
The outageAccording to Guangming Online, DeepSeek’s API service was unavailable from 8:11 to 13:04 Beijing time on October 15, about five hours. DeepSeek then issued a service-restoration and compensation notice.
The compensationAccounts get back their full API spending from October 8 to 14, a seven-day window, as credits paid automatically and drawn down first on future API billing. DeepSeek said it will thoroughly review and fix service stability, mapping and removing single points of dependency to prevent a repeat.
Tied to usageBecause refunds follow the prior week’s actual spending, heavy users get more, which is why the amounts users posted in online communities varied widely. Businesses relying on the API value a full, usage-based refund over a token gesture, but what matters most is that the outage does not recur.
▮ SIGNALAPIs are now production infrastructure for many companies, so model providers compete on uptime and accountability as well as price and capability.
❯ Reflection AI releases weights for 501-billion-parameter Beam, built to finish tasks with fewer tokens
First model, open weightsAccording to VentureBeat, Reflection AI released the weights of its first model, Beam, under Apache 2.0 on October 16, with 501 billion total parameters and 23 billion active per token. Artificial Analysis said on X that Beam scores 30 on its Intelligence Index, second among US open-weight models.
Penalizing extra tokensReflection added a penalty for unnecessary tokens during reinforcement learning while keeping the model from skipping tasks to save words. Training ran four weeks, over 100 million rollouts, on 10,500 Nvidia GB300s. The company says Beam scores about 80.1% on Terminal-Bench 2.1 using about 25,000 tokens per attempt, while GLM-5.2 scores slightly higher with about 80,000; DeepSeek V4.1 Flash scores 90.6%, ahead of Beam.
Another option for enterprisesReflection is building open frontier models and an operating system for agents, with Nvidia as an investor. The 8-bit and 4-bit versions fit on one eight-GPU server. Companies deploying their own models can do the math: shorter outputs mean the same hardware serves more requests.
▮ SIGNALOpen models are competing on a new dimension: not just who is smarter, but how many tokens it takes to get the same job done.
❯ ByteDance world-model demo stirs debate; the model is reportedly built on Seedance and targeted for release this year
A demo spreadsAccording to a TMTPost roundup, a Xiaohongshu post circulating in tech circles claims to show ByteDance’s latest world-model demo, though ByteDance has not announced it and the post names no model. A world model generates interactive virtual environments that change in real time with the user’s actions.
The reported planTMTPost cites 36Kr as saying ByteDance ranks world models first among its four AI priorities for 2026; according to people familiar, Zhang Yiming is personally overseeing the project, which builds on the video model Seedance. Citing LatePost, it says Zhang told the Seed team not to treat distillation as a shortcut, and other reports say the goal is a model rivaling Google’s Genie 3 by year-end. ByteDance has not confirmed any of this.
A new role for PicoTMTPost says the model is planned for Pico headsets, and the PICO Space Pro has been pushed to the fourth quarter. ByteDance’s Seed3D 1.0, released in October 2025, can be imported into Nvidia’s Isaac Sim. If the world model ships, Pico becomes not just a headset but a device for real-time generated content.
▮ SIGNALThe world-model race is not only about technical approach but about who already owns the content ecosystem and devices to carry it.
❯ US “Super Intelligence Force” holds first meeting, commissioning an AI risk report and early-warning rules for labs
First meeting agendaAccording to Inside Defense, citing a White House readout, the Trump administration’s Super Intelligence Force held its first meeting on October 15, commissioned a risk report from an advisory council chaired by David Sacks, and set up a working group to oversee governance commitments the White House brokered with leading AI developers.
Which risksPer the readout, topics included cyber and biosecurity, incident notification, law enforcement and consumer protection coordination, and workforce impacts. Bloomberg reports the force plans protocols for AI labs to give the government early warning of risks (Techmeme summary).
A longer compliance listOnce early warnings and incident reports become routine, US AI labs take on another ongoing compliance duty. Specific assignments, timelines and membership have not been detailed publicly.
▮ SIGNALBefore any legislation, the government is starting with “report when something goes wrong,” showing the gap it most wants to close is information.
❯ SoftBank weighs buying security firm Allied Universal at about $18 billion in equity value, eyeing robots
Early stagesAccording to Reuters via The Star, citing a source, SoftBank is considering buying US security services firm Allied Universal in a deal valuing it at about $18 billion, as Bloomberg first reported. Talks are preliminary, no final decisions have been made including on price, and SoftBank declined to comment.
What Allied doesAllied Universal supplies security staff to businesses, venues and major events and is growing in the data center industry. The deal fits Masayoshi Son’s bet on AI and robotics: SoftBank would eventually look to replace some security staff with robots, though humanoid security workers are likely further off, the source said.
A ready-made testbedSoftBank agreed last year to buy ABB’s robotics business for $5.4 billion. Buying a labor-heavy services company gives robots a ready deployment setting and customer network instead of building a market from scratch.
▮ SIGNALCapital betting on robots is starting to buy the industries that employ the most people, so the buyer sets the pace of replacing labor.
❯ Workday buys C1.ai to manage access for both employees and AI agents on one platform
Definitive agreementWorkday announced in an SEC filing on October 13 that it signed a definitive agreement to acquire C1.ai. The price was not disclosed, and the deal is expected to close in Q4 of Workday’s fiscal 2027, ending January 31, 2027, subject to customary conditions.
What C1.ai sellsC1.ai, formerly ConductorOne, does identity governance: it connects to a company’s email, CRM, code repositories, finance systems and cloud infrastructure, granting people or AI agents access when needed and revoking it afterward. Citing Deloitte, the company says only 21% of large companies have a mature way to govern agents.
From registry to controlWorkday says the deal extends its Agent System of Record from registering agents to controlling what they can access. The more agents a company runs, the more it needs one place to manage who can touch which system, the gap Workday wants to fill.
▮ SIGNALOnce agents enter the enterprise, access control becomes essential, and HR software companies want to extend from managing people to managing machines.
❯ Medical AI certifier Scarlet raises $50 million Series B led by Headline
The roundAccording to Tech.eu, London-based Scarlet raised a $50 million Series B led by Headline, with Notion Capital joining and Atomico, Creandum, Kindred Capital and Entrepreneurs First participating. Headline partner Nancy Xiao Mclaughlin joins the board.
How certification worksScarlet is a designated certification body under the EU Medical Device Regulation that reviews AI and software medical devices: customers upload technical files, reviewers assess them against regulations and issue a decision. Customers pay a subscription covering surveillance, annual audits and product changes. It says it is the only one of about 52 such bodies specializing in software and AI, holds UK MHRA designation and US FDA third-party reviewer recognition, and counts Flo Health and NHS England among customers.
Where the money goesThe co-founders built AI medical devices at Babylon Health, and the company spent three years obtaining designation before issuing a certificate. The funds will support expansion into the US and Asia-Pacific and more AI reviewers and engineers. For medical AI companies, review queues often take longer than development, so a software-focused certifier can shorten time to market.
▮ SIGNALThe faster AI medical products iterate, the more regulatory review looks like the bottleneck, and helping companies through it has become a good business.
❯ CoreWeave’s three co-founders have sold $3.3 billion of stock since the IPO as shares fall more than 60% from the peak
The salesAccording to Seoul Economic Daily, citing Bloomberg, co-founders Michael Intrator, Brian Venturo and Brannin McBee have sold $3.3 billion of stock since the March 2025 IPO, including $2.4 billion in 2026; before listing they cashed out $488 million through tender offers.
Up, then downCoreWeave rents out AI compute. Its shares rose to about 3.5 times the offering price within three months of the IPO, then fell more than 60% from that high and are down about 50% over the past year. The three still hold about 20%, and their stakes have lost about $15 billion in value since the June 2025 peak. A spokesperson called the sales routine diversification.
Credit worriesNvidia, which owns about 10%, says the risk of a CoreWeave credit event is low, and Jensen Huang said its finances are improving and its cost of capital is trending lower. Continued founder selling combined with credit-market concerns amplifies doubts about the debt-heavy model of compute rental firms.
▮ SIGNALCompute-rental stocks increasingly trade on their debt, not just on AI demand.
❯ Kioxia loses about two-thirds of its June peak market value as rate and competitive pressure mount
From first to a thirdNikkei Asia reported on October 16 that Japanese flash maker Kioxia’s market value has fallen to about a third of its peak roughly four months ago. According to Startup Fortune, Kioxia passed ¥44 trillion on June 12, briefly overtaking Toyota as Japan’s most valuable company.
Why it fellKioxia makes NAND flash, a core part of data center storage, and its shares soared on AI-driven storage demand. Nikkei Asia says rising rates threaten data centers’ business model, higher data-center borrowing costs pose a risk to memory suppliers, and Chinese competitors are closing in.
Waiting on earningsNikkei Asia notes Kioxia will report April–September results on October 30, expected to show record profit growth. Strong earnings alongside a share collapse suggest the market worries about how long the cycle lasts, not current orders.
▮ SIGNALAI-driven memory rallies come and go fast, and storage chip stocks are still valued as cyclicals in the end.
❯ FT: Alexandr Wang has Zuckerberg’s full backing and is touted as a potential Meta successor
A profileThe Financial Times published a profile of Alexandr Wang on October 16. Per a Techmeme summary, sources describe him as “an operator” who had “all” of Zuckerberg’s support during Meta’s AI overhaul and who is touted by some as Zuckerberg’s successor.
From Scale AI to MetaAccording to AI Weekly’s summary, Wang co-founded Scale AI; Meta paid $14.3 billion for 49% of it and made him chief AI officer. The FT credits him with launching the Muse Spark model, building a research group and reshaping Meta’s AI unit; he also trimmed the Superintelligence Labs team, writing in a memo that smaller teams make each person “more load-bearing.”
Betting on an outsiderThe FT’s read is that Zuckerberg bet on a founder rather than a veteran researcher, valuing an outsider’s urgency. The “successor” talk reflects others’ views; Meta has no announced handover. The piece reads more like an outside verdict on Meta’s AI overhaul.
▮ SIGNALWhen big companies rebuild AI, they increasingly favor people who lead and decide quickly over the scientists with the most papers.
❯ Apple revises App Review Guidelines to target AI nudify apps and features revealed only after approval
What changedAccording to 9to5Mac, Apple updated its App Review Guidelines on October 16. Section 1.2 on user-generated content now lists “nonconsensual intimate imagery,” and developers lacking safeguards or failing to fix violations after notice may be in breach of the Developer Code of Conduct.
Hidden features namedNew Section 2.5.19 covers hidden, dormant or undocumented features, requiring functionality to be clear to users and App Review; hiding features through code obfuscation can lead to removal and account termination. Apple is targeting apps that misrepresent their purpose in review and reveal harmful capabilities such as nudification only after approval.
End of a gray areaFor AI image apps, the gray practice of switching on features remotely after review now carries much higher risk. Section 5 also strengthens language: apps facilitating human trafficking or child exploitation will be removed and blocked from alternative distribution.
▮ SIGNALGenerative AI lets violating features hide behind a model, so app review is shifting from checking the interface to checking whether features are honestly disclosed.
❯ Apple calls iPhone Duo demand “extraordinary” as preorder deliveries slip to January
Into next year within hoursAccording to 9to5Mac, preorders for the iPhone Duo, Apple’s first foldable, opened on October 16 starting at $1,999. Estimated delivery slid from late October to November, December and then January, and to early February in some countries such as Brazil.
Apple's responseApple called demand “nothing short of extraordinary,” said it is working to get the product to customers quickly and will keep accepting online orders. Scalpers on eBay and X listed units for up to $10,000. Bloomberg’s Mark Gurman said on X that Apple launched the product before supply was ready.
What about storesApple’s statement mentions only online orders and says nothing about launch-day stock in stores. Customers without preorders may only be able to try the phone in stores for now and wait until next year to get one.
▮ SIGNALFoldables waited years for Apple, and the shortage shows the category never lacked demand, only a brand people trusted enough to buy.
❯ Tesla unveils second-generation Roadster that flips through the air on compressed-air thrusters, due in 2028
A stunt revealAccording to The Drive, Tesla unveiled the second-generation Roadster on October 15 at SpaceX’s rocket test site in McGregor, Texas. “Attitude control thrusters” running on compressed air are built into the rear diffuser, and the demo showed the car going airborne and rolling. Tesla claims 0.81 seconds from 0 to 60 mph and plans deliveries in 2028.
Suction fan and three trimsAn active aero fan pulls the car toward the ground, and Musk says it can exceed 2G in lateral cornering. According to Torque News, there are three trims: the road-going Performance with a $50,000 deposit, the limited track-focused Thrust with rear thrusters and a $250,000 deposit, and the invite-only single-seat Rocket.
Legality unresolvedAsked whether it is street legal, Musk said he was not sure it would be legal in all jurisdictions, that it “depends on your street.” Tesla’s reservation page says the thruster flight package “is not approved for public road use,” and the 0.81-second figure has no third-party test yet.
▮ SIGNALOnce sports cars borrow rocket technology, what stands in the way becomes traffic law, which never imagined such a car.
❯ A developer spends 16 days with Opus 5.5 turning The Little Prince into a 3D browser game
Planets you can exploreAccording to 36Kr, developer Sac unveiled The Little Prince: Across the Stars on X on October 13, having spent 16 days turning the book’s planets into an explorable 3D story game. It runs in the browser, requires WebGPU support, and the official site offers eight languages including Chinese, English and Japanese.
Three toolsIt was built with Opus 5.5, Tripo and Blender: Tripo generates 3D models from text or images, and Blender handles modeling, materials, animation and rendering. Players have reported bugs, such as movement keys reversing after the flower-watering task. Sac says he will open-source the whole game and publish a detailed making-of.
A lower bar for solo creatorsA 3D narrative game that once needed a small team can now reach a playable version in weeks from one person with large models and AI modeling tools. What decides quality is increasingly human judgment about subject and taste.
▮ SIGNALWhen AI drives production costs down to what one person can afford, the scarce resource is no longer technique but knowing what to make and how it should look.