Anthropic's updated IPO filing shows $33bn in revenue and a mid-November listing
# The Claude maker is about to become one of the biggest AI names on the public market.
> the_news
Anthropic filed an amended Form S-1 with the SEC on November 2, setting up an IPO in mid-November. It plans to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The filing reports $33.24 billion in revenue for January through September, up roughly 1,200 percent on the same stretch of 2025, and lays out heavy spending on AI chips and rented compute. The company has not yet named the number of shares or a price range. Its most recent private round valued it at $965 billion, which would make this one of the largest tech listings in years.
> why_it_matters
An Anthropic IPO puts a frontier AI lab's real finances in front of public investors for the first time, revenue, losses and the enormous cost of compute alike. How the market prices ANT will shape how every other AI company thinks about going public.
> real_life_use_case
A founder planning a raise can read Anthropic's S-1 as a live benchmark, how investors weigh breakneck revenue growth against billions in compute costs, and use those same framings when pitching their own AI margins and burn.
Moonshot releases Kimi K3.1 Pro open weights, a 2.8-trillion-parameter model
# An open model is now trading blows with the strongest closed systems on coding and agents.
> the_news
Moonshot AI released the open weights for Kimi K3.1 Pro on November 4. It is a 2.8-trillion-parameter mixture-of-experts model that activates 104 billion parameters per token and handles a one-million-token context window. The model adds sparse attention to every full-attention layer for efficiency and ships with a vision encoder. On reported benchmarks it scores 65.7 on Terminal-Bench 4.0 and 94.2 on GPQA-Diamond, matching strong closed models on coding and agentic work. The weights are on Hugging Face under a custom Kimi license that restricts some commercial use, and inference engine vLLM added day-zero support across several chip vendors.
> why_it_matters
A top-tier open-weight model that teams can download and run themselves keeps pressure on closed labs and gives builders a serious alternative they can host, inspect and fine-tune, without being locked into a single vendor's API and pricing.
> real_life_use_case
A team building a coding agent can self-host Kimi K3.1 Pro to keep sensitive code in-house, tune it on their own repositories, and avoid per-call API fees, while still getting near-frontier results on real software tasks.
Tencent launches Hy4, a 770-billion-parameter model that can see screens and use software
# The model doesn't just answer questions, it operates the apps on your desktop.
> the_news
Tencent launched Hy4 on November 4, adding visual understanding and computer-use skills to its flagship AI model. The company says Hy4 can read charts and documents, reason across several images at once, and drive both command-line tools and graphical interfaces, with targeted training on applications like Blender and CAD software. The 770-billion-parameter model is available across Tencent's products and on third-party platforms including OpenRouter, with weights and a system card promised soon. API pricing stays at $0.834 per million input tokens and $2.501 per million output tokens, with a 50 percent discount on Tencent Cloud through December 31.
> why_it_matters
Computer-use models that can actually click buttons and run desktop software move AI from a chat box toward doing real multi-step work inside the tools people already use. Native support for design and engineering apps points at where this is heading.
> real_life_use_case
A design studio could hand Hy4 a brief and let it set up a scene in Blender or draft a layout in CAD software, handling the repetitive clicks while a human reviews and refines, instead of doing every step by hand.
Ex-Google chip researchers release Ripple-65B, a small open model that punches above its size on coding
# Sixty-five billion parameters beating models more than ten times larger on key coding tests.
> the_news
Startup Ricursive Intelligence released Ripple-65B Preview on November 4, an Apache 2.0 open-weight model its creators say leads U.S. open models on two coding benchmarks, Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard. Founders Anna Goldie and Azalia Mirhoseini previously helped build Google's AlphaChip. Ripple is a dense 65-billion-parameter model trained from scratch on 12 trillion tokens, and the company says the full training effort cost about $32 million, a fraction of frontier budgets. It is on Hugging Face and via partners including Together AI and OpenRouter. The company is candid that its lead is narrow and some rival models score higher on other tests.
> why_it_matters
A compact, cheaply trained model that competes with giants challenges the assumption that only massive budgets and parameter counts win at coding. Smaller open models are easier to run locally and cheaper to serve at scale.
> real_life_use_case
A startup can download Ripple-65B and run it on a single workstation or modest GPU for its internal coding agent, getting strong terminal and software-engineering performance without paying for a frontier API or renting a cluster.
Nvidia will fill only about 1% of China's RTX Pro 5500 chip orders
# Chinese firms ordered six million chips. They'll get roughly sixty thousand.
> the_news
Nvidia has decided to fill only about 1 percent of the roughly six million RTX Pro 5500 chips ordered over four months by Chinese AI and cloud companies including Alibaba, Tencent and ByteDance, Bloomberg reported on November 4. At over $11,000 each, the orders are worth an estimated $70 billion, but buyers will receive only tens of thousands of units. The RTX Pro 5500 is not powerful enough to fall under U.S. export restrictions, yet Nvidia is prioritizing higher-margin chips for customers like OpenAI and Anthropic while global production stays tight and capacity shifts to its next-generation Rubin line.
> why_it_matters
Compute supply, not model design, is becoming the real constraint on who can build AI at scale. When even a mid-tier chip is this scarce, access to hardware starts to decide which companies and countries can keep training and serving large models.
> real_life_use_case
A team planning an AI product should treat GPU access as a scheduling risk, locking in cloud capacity or reserved instances early and designing to run on smaller or open models, rather than assuming the chips they want will be available on demand.
OpenAI says ChatGPT now runs search queries at one-fifth of Google's volume
# "Ask Chat" is quietly becoming the new "Google it."
> the_news
OpenAI said the volume of search-style queries run through ChatGPT now equals about 20 percent of the volume run through Google. Adam Fry, OpenAI's consumer product lead for ChatGPT, shared the figure at a press event on November 3, defining a search as any query that triggers a web search rather than a simple definition or how-to. For comparison, Microsoft's Bing sat around 3 percent of the market after more than a decade. OpenAI is also leaning into ads and what it calls "Intelligent UI," building visual, interactive answers inside chat responses as it eyes a possible IPO.
> why_it_matters
Search is how most people find information and how a huge share of ad money flows online. If a chatbot is already handling a fifth of that volume, the habits and economics that Google built over two decades are genuinely in play.
> real_life_use_case
A marketer or founder who relies on being found should start optimizing for AI answers, clear, citable content and structured facts, not just classic SEO, because a growing share of their audience is asking ChatGPT instead of typing into Google.
Google DeepMind gives its Antigravity agent chemistry skills that cut drug-discovery loops to an hour
# Work that took medicinal chemists days now runs while you grab a coffee.
> the_news
Google DeepMind introduced Cheminformatics Skills for its Antigravity agent workbench on November 3. In a lung-cancer example built around the drugs crizotinib and lorlatinib, an agent searched the literature for thousands of compounds, ran matched molecular pair analysis, trained potency models and proposed new candidate molecules with synthetic routes, all in about an hour, work that would normally take days. The skills split the job deliberately: the language model orchestrates and reasons while deterministic tools grounded in the open-source RDKit toolkit do the exact chemistry. DeepMind is also offering 10,000 scientists six months of free access to Antigravity's top tier.
> why_it_matters
This shows a useful pattern for serious science work: let the AI plan and read the literature, but hand the precise calculations to trusted deterministic tools. DeepMind is clear that every computational design still has to be synthesized and tested in a lab.
> real_life_use_case
A pharma or biotech team can use Antigravity's chemistry skills to screen and shortlist candidate molecules in hours instead of days, freeing bench chemists to spend their time synthesizing and validating the most promising designs.
Google draws backlash for tying school Search access to turning on its AI
# Want Search for your students? You'll have to switch on the AI too.
> the_news
Google announced on November 3 that Gemini, Gemini Notebook and Google Search will all be off by default for K-12 student accounts, with administrators choosing which to enable. The catch drawing criticism: Google counts Search as an AI product, so a school that turns Search on gets AI Mode and AI Overviews with it, with no option for plain search without AI. Critics including Senator Bernie Sanders called it coercive. Google says the changes give schools more control, not less, and that it is adding a "Guided Learning" mode meant to tutor students through problems rather than hand over answers.
> why_it_matters
How AI reaches classrooms is becoming a policy fight, not just a product decision. Bundling search and AI together forces schools into an all-or-nothing choice and raises real questions about control, consent and what tools students are nudged toward.
> real_life_use_case
A school IT administrator weighing these settings should map out exactly which services each toggle enables before rolling them out, and prepare an alternative search option for classrooms that want research tools without generative AI in the results.
Nvidia unveils DSX POD modular AI factories and commits roughly $10bn to Europe
# Prefab data-center containers that drop AI capacity in two weeks instead of six months.
> the_news
At its Berlin keynote on November 4, Nvidia unveiled DSX POD, a reference design for modular AI factories built on its Vera Rubin systems. Partners manufacture pre-integrated, factory-validated container units, cutting on-site deployment from six months to 14 days once utilities are ready. Nvidia also outlined a DSX AI Factory Platform it says can cut cooling water use sharply and deliver the same output with up to 30 percent less power. The company committed nearly $10 billion to Europe's AI ecosystem and introduced an OpenShell reference architecture for safer AI agents, plus industry-specific open models.
> why_it_matters
The bottleneck in AI right now is physical, power, cooling and how fast you can stand up data centers. Standardized, prefab designs that slash build time and energy use directly affect how quickly and widely AI capacity can actually be deployed.
> real_life_use_case
A company or government planning its own AI infrastructure can adopt a validated modular design like DSX POD to go live in weeks rather than months, and lean on the efficiency gains to keep power and water costs manageable as it scales.
webAI releases TwIL-8B, an 8.8-billion-parameter model that tops its class on math
# A model small enough to run on a laptop, scoring 98 on a hard math benchmark.
> the_news
webAI released TwIL-8B on November 4, an 8.8-billion-parameter reasoning model aimed at math, code and instruction-following. Post-trained from IBM's Granite 4.2 8B, it leads its size class on math and code, scoring 98.2 on MATH-500, the top mark in webAI's own comparison, though it trails larger models on broad knowledge tests like GPQA Diamond. The pitch is cost per task rather than raw scale: the 4-bit build is under 5 GB, small enough to run on a CPU or a modest GPU. Weights are on Hugging Face under a non-commercial license.
> why_it_matters
Small models that run locally and still handle real reasoning make AI cheaper, more private and usable offline. For many focused jobs, an efficient 8-billion-parameter model beats paying for a giant frontier model you don't need.
> real_life_use_case
A developer building a math or coding helper can run TwIL-8B entirely on a laptop or small server, keeping data local and costs near zero, while reserving a larger cloud model only for the occasional broad-knowledge query it can't handle.