Daily AI News — 22 Oct 2026

DeepSeek has raised roughly 100 billion yuan, about $15 billion, in its latest funding round, Bloomberg reported on 22 October, citing people familiar with the matter. The Hangzhou company lifted its target twice during talks with investors, from 50 billion yuan to 80 billion before landing on the final figure. Battery maker CATL and Tencent were among the biggest backers. The raise is one of China's largest ever for a startup and signals DeepSeek's intent to compete globally with OpenAI and Anthropic, both of which have each pulled in more than $100 billion. DeepSeek still has far less cash, but the gap narrowed sharply.

DeepSeek built a reputation on doing more with less. Handing it $15 billion changes the calculus: it can now buy serious compute and talent while keeping its efficiency edge. That makes the US-versus-China model race closer and keeps downward pressure on prices across the whole market.

Builders weighing where to spend their token budget get another well-funded frontier option. A team can benchmark DeepSeek's models against OpenAI and Anthropic on their own workloads and route traffic to whichever gives the best cost-to-quality ratio.

Anthropic introduced Claude Duet on 21 October, its answer to Meta's Muse and OpenAI's Dots. Duet is a proactive agent built on Opus 5.5 that connects to your apps, sends a short morning brief, checks in every few hours, and alerts you when something needs attention. It runs in the cloud with its own virtual machine and browser, so it can complete tasks without touching your computer. It asks for approval before sending a message, making a payment, signing a document, or deleting a file. Duet is in research preview via waitlist, limited to US Pro and Max subscribers on web and an iOS app.

Agents so far have mostly answered when spoken to. Duet, Dots, and Muse push the model toward noticing and acting first, which is where assistants start to feel genuinely useful for busy people. The approval gates on payments and messages also show the labs taking autonomy risk seriously.

A small-business owner can let Duet scan email, calendar, and project tools overnight, then wake up to one brief flagging the invoice that needs chasing and the meeting that moved, instead of digging through five apps themselves.

At GTC Berlin on 21 October, Nvidia unveiled DSX POD, a reference design that lets partners build complete Vera Rubin AI factories inside shipping containers, assembled and tested at the factory before delivery. Nvidia says it can cut deployment at a prepared site from six months to as little as 14 days. Dell, HPE, Foxconn, Schneider Electric, and Vertiv plan worldwide deliveries from early 2027. Nvidia also launched the OpenShell Reference Architecture for running AI agents inside hardware-enforced boundaries, where its BlueField-4 chips can stop an agent in milliseconds if it strays. Huang said the ChatGPT moment for robotics is fast approaching.

Compute is the real constraint on AI right now, and the bottleneck has moved from chips to how fast you can stand up a data center. Shrinking that from months to weeks reshapes who can deploy at scale and how quickly, while OpenShell bakes agent safety into the silicon itself.

An enterprise or sovereign-cloud operator can order a pre-integrated container, connect power and cooling, and be producing tokens in weeks rather than half a year, then run its agents inside OpenShell so a misbehaving one is contained by hardware, not just policy.

Walmart on 21 October introduced tools that let AI agents shop and complete transactions for customers. A new Walmart Agent Gateway lets assistants such as OpenAI's ChatGPT, Google's Gemini, and Meta's Muse connect directly to Walmart and Sam's Club systems to search products and buy. Next month Walmart adds Agent Profiles, letting shoppers authorize individual agents, track their purchases separately, set spending limits, require approval per transaction, and revoke access anytime. Agents are barred from payment card data and can be blocked for breaking rules. The move contrasts sharply with Amazon, which has sought to limit third-party AI shopping agents on its platform.

Agentic commerce has been mostly theory. A retailer Walmart's size wiring agents directly into checkout makes it real and sets an early standard for how AI buys on our behalf. The split with Amazon also marks a genuine strategic fork over who controls the agent-to-store relationship.

A shopper can tell their assistant to reorder household staples under a monthly cap, then review a separate agent order log and pull authorization instantly if a purchase looks wrong, keeping control while offloading the chore.

At its GM Forward event in New York on 21 October, General Motors said it will use AI and new software to cut vehicle development time roughly in half, to about two years, to better compete with fast-moving Chinese automakers. CEO Mary Barra also unveiled the OnStar AI Assistant, a proprietary voice agent, activated by "Hey OnStar," that answers questions about vehicle alerts, charge range, and features using real-time vehicle data. It debuts on the 2027 Cadillac Lyriq before rolling out over the air, giving customers an alternative to Google Gemini. GM touted roughly 3 billion virtual training miles per day toward eyes-off highway driving by 2028.

Development speed is where Chinese automakers have pulled ahead. If AI genuinely halves GM's cycle, it changes how fast legacy carmakers can respond to the market. The OnStar agent also shows a 30-year-old service reinventing itself from a call center into an in-car intelligence platform.

A GM owner can ask the OnStar agent why a dashboard light is on or whether the battery will reach the next charger, and get a plain-language answer from live vehicle data, instead of guessing or calling a dealer.

Tesla has begun production of its Optimus humanoid robot in California, the company said on 21 October, even as quarterly profit fell 43%. The move marks a shift from prototypes and staged demos toward building robots at volume, part of CEO Elon Musk's long-stated plan to make humanoids a core Tesla business. Reporting has described Tesla ramping Optimus output at its Fremont factory from a few dozen units weekly earlier in 2026 toward several hundred, with the facility pivoting away from Model S and Model X production. Tesla frames the robot as a future growth engine even as its core car business faces pressure.

Humanoid robots have been perpetually almost here. A major manufacturer actually starting a production line, rather than showing a stage demo, is a concrete step. It also signals Tesla leaning on robotics to offset a shrinking car business, a bet that will shape the whole sector's credibility.

Factory and warehouse operators watching the humanoid market get a real data point on cost and availability. Early adopters can start scoping repetitive physical tasks, sorting, moving, loading, that a general-purpose robot might take on before committing to fleets.

AI startup Recursive released Flint on 21 October, a free coding model small enough to run on a laptop that the company says was built largely by its own automated research system, the Eureka Machine. A small team built it from scratch in under four months. Flint is a mixture-of-experts model using 2.5 billion of its 40 billion parameters at a time, and it is open under Apache 2.0 on Hugging Face. In tests it generated about 131 tokens per second on an M4 Pro MacBook and beat other sub-50-billion-parameter models on the SWE-Bench Verified bug-fixing benchmark. Recursive, led by former Salesforce chief scientist Richard Socher, emerged from stealth in May with $650 million.

Flint is an early, concrete test of recursive self-improvement, the much-debated idea that AI can build better AI. It is deliberately modest and human-overseen, but a capable model shaped mostly by an automated research loop is exactly the kind of evidence that moves this argument from theory to practice.

A developer who wants privacy or no usage fees can download Flint and run a capable coding assistant entirely offline on a laptop, keeping source code local while still getting strong help on real bug fixes.

Swedish legal AI company Legora announced a $550 million Series E on 21 October, valuing it at $8.8 billion and bringing total investment past $1.4 billion. Atreides Management led the round, with new backers including Fidelity, Capital Group, J.P. Morgan, Baron Capital, and Glade Brook, alongside existing investors like Accel and Benchmark. Legora says annual recurring revenue has passed $200 million, after going from $1 million to $100 million in under 18 months, then adding another $100 million in under six. Its agentic platform, Legora aOS, helps more than 140,000 legal professionals across 2,400 firms research, track legislation, and handle requests. The raise comes six weeks after rival Harvey's $550 million round.

Legal work is text-heavy, high-stakes, and expensive, which makes it one of the clearest proving grounds for AI that does real professional work rather than just chatting. Two large raises within six weeks show investors betting that AI is reshaping how law firms are staffed and how their work gets done.

An in-house legal team can lean on Legora's agents to run first-pass research, flag relevant legislative changes, and triage incoming requests, freeing lawyers for the judgment calls and advice that actually need a human.

Nvidia said on 21 October it is expanding healthcare and life-sciences work around protein design, autonomous labs, and adaptive radiotherapy. With the University of Washington's Institute for Protein Design, it introduced RFD4-Proteina, a new open BioNeMo model that combines programmable protein design with search, letting scientists and agents generate and refine candidates using molecular inputs, constraints, and scoring objectives. Nvidia said it was tested across four benchmark suites, protein binding, small-molecule binding, enzyme scaffolding, and DNA binding, and beat the baselines on design success rates and throughput. The model is available through partners including Amazon, Anthropic's Claude Science, Google Cloud, Hugging Face, and OpenAI. Nvidia is also working with Siemens Healthineers on AI-assisted cancer treatment planning.

Designing a protein that binds a chosen target is central to new medicines and materials, and doing it computationally with higher success rates compresses a slow, costly step. Making the model open and available through major cloud and AI partners puts that capability within reach of far more labs.

A biotech researcher can describe a target and constraints, have RFD4-Proteina generate candidate protein designs, then score and refine the shortlist in software before spending money on wet-lab synthesis, cutting both time and cost.

Meta said on 21 October that its Muse AI assistant has surpassed 10 million users, with shopping emerging as a top activity. The milestone lands as Meta pushes Muse across its apps and devices and positions it against OpenAI's and Anthropic's proactive agents. Shopping's prominence fits the day's broader theme, Walmart opened its systems to agents including Muse, and points to where consumer AI is finding real daily use. The disclosure came alongside other Meta news, including a plan to fund its heavy AI buildout through a newly created class of stock, underscoring how aggressively the company is spending to compete in AI.

User counts are noisy, but 10 million with a clear habit forming around shopping tells you where consumer AI is actually sticking. It also sets up a direct contest among Muse, ChatGPT, and Claude for the everyday assistant slot, and ties Meta's assistant into the emerging agentic-commerce stack.

A shopper already in Meta's apps can ask Muse to find a product, compare options, and, through tie-ins like Walmart's Agent Gateway, move toward checkout without leaving the chat, turning a casual assistant into a buying tool.