The newest item sits on top. Tags are there for chips, agents, or the data center arms race.
Corrections: if a fact is wrong, email hello@loadedtokens.com.
Categories
Bloomberg reports that Suzhou Denglin Technology and Jiangsu Xinde Semiconductor Technology were due to start trading in Hong Kong on Tuesday, after their IPOs raised a combined $301 million. According to Bloomberg, the listings arrive while tech stocks are swinging sharply, so the debuts will test how much investors want chip shares.
Semafor reports that lawmakers are looking at Nvidia's chief executive after a spy scandal. According to the outlet, they are considering tighter export controls that would restrict the company's chip sales to Beijing, along with possible investigations. The excerpt does not describe the scandal itself.
TechCrunch reports that Nvidia's stock fell more than 4% after its board began investigating Jensen Huang's relationship with a person suspected of spying for China. A securities law expert quoted by the outlet warned that Nvidia's earlier claim that the matter was "normal" could prove costly if shareholders sue.
Bloomberg reports that Nvidia CEO Jensen Huang had an affair with a person alleged to be a Chinese spy. According to the outlet, members of both parties in Congress are now pressing for more information. They want to gauge what the matter could mean for national security. The excerpt describes allegations and early demands, not findings or any action taken.
Bloomberg reports that Nvidia's general counsel got involved in a US government investigation into suspected AI chip smuggling by Singapore-based Megaspeed International. The lawyer reportedly asked whether the Trump administration had a hand in detaining a Megaspeed executive. Bloomberg says that executive is now central to a scandal that threatens Jensen Huang.
LangChain has a new post arguing that agents burn time and context on clumsy searches, oversized results and outdated data. The company describes the retrieval techniques its Deep Agents harness uses to give agents the right context. Builders working on agent memory may find it a useful practical reference.
Bloomberg reports that Google is releasing a new set of AI chips to build on its rise as a competitor to Nvidia. The excerpt is a short video segment featuring Bloomberg Intelligence analyst Mandeep Singh. It gives no chip names, specs, pricing or availability, so the business details still need confirming.
Bloomberg reports that Fireworks AI, an Nvidia-backed startup that runs AI models for other companies, told customers on Monday that attackers had accessed their data. According to Bloomberg, this is the company's second known hack in a few weeks. The excerpt doesn't say what data was exposed or how many customers were affected.
The Guardian reports that Nvidia's board must make a difficult decision about Jensen Huang after the co-founder had an affair with an alleged Chinese spy. Huang has led the chipmaker since he co-founded it in 1993, and the outlet notes his children hold leadership roles there. The piece asks whether he could be pushed out. The excerpt does not report a board decision.
AWS says Amazon Bedrock Managed Knowledge Base now handles single-shot RAG and retrieval-only requests through AgenticRetrieveStream, and agentic retrieval is still available. AWS says a parameter called maxAgentIteration controls how much retrieval effort each request gets. Its walkthrough uses a made-up internal banking policy assistant as the example.
AWS says NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter vision-language model, can now be deployed from SageMaker JumpStart. Teams can choose an NVFP4 or BF16 variant and run inference on text, images, and video. AWS is positioning it for document understanding, computer-use agents, and other multimodal agent work.
In a press release, Reindeer says its enterprise AI agents can now be bought by Google Cloud customers through the Google Cloud Marketplace. The company describes the agents as built on Gemini models and deployed on Google Cloud, and it calls the move a wider partnership aimed at core business operations. This is a vendor announcement, and the excerpt gives no customer results or pricing.
The Financial Times reports that US stock indices bounced back after a sell-off tied to AI. Lower oil prices helped the broader market look past a decline in Nvidia shares. The excerpt does not give figures or explain what started the AI rout. That makes this a read on investor mood, not news about the technology.
Anthropic's Claude Code release notes say Claude Fable 5.2 is now the default Fable model, with a 1M-token context window. It is priced at $8 input and $40 output per million tokens, and cache reads cost $0.20. The update also fixes a bug where MCP OAuth could pass refresh tokens to a different host after a server redirect.
The Verge reports that Logic Pro now has a "Buy More Usage" option, a one-time $49.99 purchase. It extends access to Apple Intelligence features that carry usage limits. Apple first described those limits in a support article in early September. According to the report, 9to5Mac spotted the change, and Logic Pro includes a built-in AI assistant that can help with tasks such as starting a new song.
OpenAI's status page lists an incident with elevated errors in ChatGPT conversations, now marked resolved. The company says every affected service is back to normal and the Conversations component shows as operational. The post does not give a cause, duration, or how many users were hit.
Bloomberg reports that Anthropic has released updated versions of its Mythos and Fable AI models. According to the outlet, the launch is meant to strengthen Anthropic's standing as a leading AI developer just before the company goes public. The excerpt doesn't describe what's new in the models, how they perform, or when they'll be available.
The Verge reports that Anthropic has released two models, Claude Fable 5.2 and Claude Mythos 5.2, and is positioning them for hard technical work. Anthropic says both improve on the 5.1 versions in coding, reasoning, and vision, and also produce output faster. These are the company's own claims. The report does not include any independent testing.
Bloomberg reports that Ericsson's new chief executive expects artificial intelligence to boost demand for next-generation telecom networks. According to the report, the reasoning is that AI-connected gadgets such as smart glasses will need more data. For now this is an executive's outlook and does not include figures or a timeline.
Data Center Dynamics reports that Digital Realty has opened a liquid cooling lab in São Paulo, Brazil. According to the outlet, it is the company's first DRIL lab in Latin America. It comes after a London lab that opened earlier this year.
Anthropic's engineering post describes how Claude chooses what to remember, including a dreaming pass that organizes memory and corrects mistakes. It explains the product. It does not hand you a new setting. You cannot schedule the dream from the blog post.
OpenAI and Ligo published BioBasicsBench, 100 tasks about raw wet-lab evidence, including the times the data cannot support a conclusion. It scores agents. It does not hand you a lab partner.
You can talk to a Cursor agent while it works, on the desktop app, the web, and iOS. The call runs on Grok Voice. Dictation is still the quieter cousin: it only types into the box.
OpenAI will work with Hacktron to hunt bugs in OpenAI's own systems, especially where those systems touch third parties. Hacktron also gets broader Daybreak Red access. This is a security program. There is no switch in your settings.
A November 6 changelog note lists a fully local Computer on Mac, plus a voice mode next to Teammates. The note does not spell out plan limits. Read it before you promise either feature in Slack.
Web Answer on Bedrock AgentCore is generally available and returns a cited answer over MCP. It is billed per call and, for now, only in US East (N. Virginia). If your stack lives somewhere else, this is a postcard.
Claude Console organizations can set an IP allowlist, in beta, with up to 1,000 addresses or ranges. Requests from anywhere else are refused, which limits what a leaked key can reach. It does not replace rotating the key. Change the locks too.
Claude Enterprise admins can sign a member out of Claude everywhere, in beta, through the Admin API. The person stays in the organization and can sign in again. It is an off switch for sessions. Their name stays on the company list.
A cloud agent can scaffold an app, add Postgres and Cursor sign-in, and deploy it after security checks. You need a paid plan with cloud agents. Nothing goes live until you explicitly publish.
ERNIE 5.5 Flash Base is a 280-billion-parameter mixture-of-experts checkpoint, 14 billion active, stopped before instruction tuning. The license does not cover the EU, the UK, or South Korea. Check the map before the model card gets you excited.
Gemini Omni 1.2 Flash makes native clips up to 15 seconds. Google priced 720p at $0.07 a second, 30% under Omni 1.1. It is generally available on the Gemini API and Google Cloud, and it stamps a SynthID watermark.
OpenAI's new cybersecurity model is rated Critical under its Preparedness Framework and starts with approved Daybreak Red customers. It finds and fixes vulnerabilities inside an authorized scope. The all-hands email is still your job.
Thinking Machines' new model has 594 billion parameters, 17 billion active, and a score of 31 on the Artificial Analysis Intelligence Index. Weights stay closed until outside safety testing. Apache 2.0 is the plan, not today's license.
NTT DATA said on November 5 it will deploy 900 megawatts of Bloom Energy fuel cells across data center projects, with first shipments planned for 2027. The release calls this Bloom's largest international project, and says the cells make electricity without combustion. Bloom flags the megawatts, the start date, and any expansion as forward-looking. Nothing in the release is a site you can tour.
OpenAI published community commitments for the places it builds infrastructure, plus policy proposals it says would spread the benefits. Read it as a company position. It will not pick your next API.
Gradium's Phonon runs without a server, in about 65 MB and 75 MB sizes, across five languages. The English 31-million-parameter model is Apache 2.0. The other models are non-commercial. Local does not mean "sell it in your app."
NVIDIA's toolkit release adds general-availability support for the Vera Rubin architecture and extends Windows on Arm to DGX Station. If you already build on CUDA, this is the upgrade note, not a new app for your lunch break.
Google published a health-AI evaluation framework written with outside clinicians. It looks at accuracy, safety, and usefulness, using benchmarks, autoraters, and expert review. The post shares the method. It does not publish a score for a model you can buy.
Images 2.6 is in ChatGPT, Codex, and the API, with GIF output and generation OpenAI says is 50% faster than before. It is available on every ChatGPT subscription tier. Your mascot can wave. A person still has to approve it.
OpenAI says a research pipeline produced working local privilege-escalation exploits for 948 Linux kernel issues and drafted patches. The post puts that run at $1,213,430 in API credits. It is a controlled lab result. You cannot point this at the office server.
OpenAI says it will acquire Pierre Computer, maker of Code Storage and the Trees and Diffs libraries already used in Codex. The deal has not closed. The companies stay separate until it does. Your repo does not move houses today.
Teammates are personal agents for work that runs past one prompt, and they keep project memory across runs. Perplexity is launching them on the web for Max subscribers, with iOS listed as later. Not on Max? You are watching the trailer.
As of November 3, text from Claude Mythos 5, Opus 4.7, and Sonnet 4.6 carries Anthropic's text watermark on the platforms named in its marking guide. On Sonnet 4.6, a non-default temperature, top_p, or top_k may skip the mark.
Ai2's StateBench checks whether a model can track changing facts across a long document, at four levels of difficulty. Frontier models handle the easy levels and break down on the hardest long ones. It is a benchmark. It will not file your contracts.
Message threads are in public beta on the Claude Platform. After the first request, later turns send only the new messages and the previous response id. A thread expires five minutes after the last request. Anthropic says token use stays the same.
Anthropic open-sourced FlashProve, a research prototype that proves a reply came from a committed set of weights. It is Apache 2.0 research code, not a Claude setting. Proving costs many times a normal answer. Nice receipt. Expensive printer.
LifeVal is OpenAI's evaluation of everyday tasks, from childcare and budgets to government paperwork. The post explains how the task list was built. It is a measurement project. You cannot switch it on in ChatGPT.
OpenAI raised Security Bug Bounty payouts on November 2. A critical finding in a core product such as ChatGPT, Codex, or the API now pays up to $30,000. Exceptional cases can pay more once bonuses are included. This is a researcher program, not a product setting.
On November 2, OpenAI launched Grid Protectors, a plan for future compute campuses to cut electricity use when the local grid is strained. The Effingham County agreement with Georgia Power includes up to 1,000 megawatts of flexible demand response. OpenAI compares that with the average demand of roughly 700,000 households. The campus is still in development. A pledge is not a building with the lights on.
Outbound SIP calling is now available to every organization on the Live API. You connect GPT-Live to a number through your own SIP provider. The changelog is the availability note. The price sits on a separate pricing page. You still bring the phone plan.
Microsoft published Mage-VL-Cowork, a 4-billion-parameter agent that mixes on-screen actions with a terminal, under Apache 2.0. It reports 66.0% on OSWorld-Verified with a 50-step budget. That figure is theirs. We did not sit it at a laptop.
GeekWire says an SEC filing shows Powerlattice raised $86 million and is seeking $111 million, with 18 investors so far. The filing does not name the round, and the company declined to comment. Powerlattice builds a power-delivery chiplet and says a shorter path to the processor can cut energy use by more than half. That cut is the company's claim. We did not open the filing or run the chip.
SiliconANGLE reported on October 28 that FuriosaAI raised $620 million to scale its inference chips. The company put the valuation at $2.9 billion, and South Korea's National Growth Fund put in $280 million of the round. The article says the chips use less than half the power of Nvidia's GPUs. That comparison is not a test we ran. A bigger round is not a bigger power line.
Google calls Carbon its strongest model and is starting with a subset of cyber defenders, then trusted partners in the Fairwind program. Introductory API pricing is $2 in and $10 out per million tokens. A general release is not dated.
DeepMind's Genie 4 generates interactive worlds with native audio. The rollout is a research preview for trusted testers, plus a waitlist. You cannot put this on a work project this week. The trailer is not the product.
OpenAI's fast workhorse is in the API at $1 per million input tokens and $5 per million output. OpenAI says that is 50% below GPT-5.6 Terra and GPT-6.1 Sol. Paid ChatGPT Work and Codex plans can use it too.
Nothing in this category yet.
RSSAtomJSONnewsfeed.jsonThis weekIssue archive
Tuesday's briefing is Friday 6:00 AM through Monday 11:59 PM Central. Friday's is Tuesday 6:00 AM through Thursday 11:59 PM Central. We lock the list at 8:00 PM Central the night before.
The source comes with the story.
Your inbox can handle one more.
