Mind-Reading Without Surgery and the Billion-Dollar Bid to Stop AI Financial Bleeding

Yesterday in AI  |  July 1, 2026

Mind-Reading Without Surgery and the Billion-Dollar Bid to Stop AI Financial Bleeding

The physical interface between humans, hardware, and enterprise AI models is undergoing a massive shift. This episode covers Meta's newly published Brain2Qwerty v2 research, a non-invasive brain-computer interface capable of decoding skull-external typing signals at a record 61% accuracy rate. 

We look into China's massive domestic chip push as food delivery giant Meituan trains its 1.6 trillion parameter LongCat-2.0 coding model entirely on non-Nvidia processors. We analyze Etched closing a total of $800 million in funding to build TSMC-manufactured chips dedicated entirely to cutting the massive financial costs of model inference. Plus, we break down Anthropic's new verticalized Claude Science workbench, AWS committing $1 billion to deploy engineering pods on-site to build corporate moats, and Singapore-based Acti putting Google Gemini agents straight into your smartphone keyboard.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, and welcome back. This is Yesterday in AI, your daily digest of everything happening in the world of AI in ten minutes or less. I'm Mike Robinson. It's Wednesday, July 1st, and AI is getting closer to you physically. Reading messy brain signals, living inside your smartphone keyboard, sending engineering pods directly to your office, and funding a massive hardware stack built specifically to run cheaper. Let's get into it. Meta published research detailing a system called Brain2QWERTY V2, an AI that decodes complex brain activity into type sentences without requiring surgery. A participant wears an MEG scanner to track magnetoencephalography while typing normally. As they type, the skull-mounted hardware picks up the faint magnetic fields generated by the brain, and the model reads those raw signals to construct text. There is no implant, there is no scalpel. Elon Musk's neuralink depends on physical electrodes drilled directly into the skull. Implants deliver clear results, but they demand neurosurgery and carry infection risks. Meta is capturing signals from outside the bone entirely. That data is incredibly hard to decode because magnetic fields get weak and messy when passing through bone and tissue. Brain to Courtney V2 hits 61% word accuracy. That is a massive jump over the 8% benchmark held by prior non-invasive tools. The development team targeted the tool at people who cannot physically type due to paralysis or motor neuron conditions like ALS. If a patient can still mentally mimic the hand motion, the software translates that thought into text. The clinical setting is the starting point, but the product roadmap stretches much further. Meta has hundreds of millions of hardware users. They are actively shipping smart glasses, and they have funded this specific neural research lab for years. Mark Zuckerberg is building the foundation for a post-smartphone interface, and he just bolted a neural translator onto his hardware division. While Meta is reading brainwaves using consumer hardware designs, a food delivery giant in China is showing that domestic hardware can train massive software engines under heavy US trade sanctions. Meituan is China's absolute heavyweight in food delivery and local logistics. They just announced a 1.6 trillion parameter open source coding model called Longcat 2.0. The firm built and trained the engine from scratch on a computing cluster packed with 50,000 processors made entirely inside China. The cluster contains zero NVIDIA silicon. Whitehouse trade restrictions were engineered to starve China of high-end graphics cards, hoping to block them from training massive models at scale. Maituan just proved a company can run a 1.6 trillion parameter workload using domestic silicon. Independent benchmarks are not public yet, and Maituan hid the exact name of the chip foundry in their public statements, meaning tech teams are still verifying the performance. The sheer scale of the cluster proves that Chinese state-backed hardware is much further along than trade teams in Washington want to admit. Longcat 2.0 ships with a 1 million token context window, meaning the model can hold roughly 750,000 words in its working memory during a single prompt. That is the equivalent of analyzing eight full novels simultaneously. For software developers needing to audit an entire corporate repository at once, that capacity is a massive asset. They are open sourcing the code, letting anyone download and run the engine. This massive training run shows how much money is flooding into hardware, but a domestic startup just secured a billion dollars to solve the massive financial drain of running these models once they're built. A chip design startup named Etched just confirmed it has secured $800 million in total funding. That includes a $500 million investment round that closed quietly in December 2025 at a $5 billion valuation. The company also locked down $1 billion in contracted orders for its specialized processors, which will be manufactured by TSMC. The tech sector divides computing workloads into two distinct phases. Training is the heavy lifting required to build a model, which happens once and consumes massive power. Inference is the ongoing cost of running the model every single time a user types a prompt into ChatGPT, an enterprise system pulls a recommendation, or an automated agent runs a script. Inference happens billions of times a day, and it is rapidly chewing up the operational budgets of enterprise tech departments. NVIDIA graphic cards dominate the market because they excel at training, but they are not purpose-built for ongoing execution. Etched is building specialized chips dedicated strictly to inference. Their design promises massive efficiency gains, reducing the power drain per watt during execution. The $1 billion order backlog proves that enterprise buyers are desperate to slash their recurring compute invoices. The initial hardware race prioritized model size, but this next phase is about controlling the financial bleeding of live operations. While hardware makers are rebuilding chips to cut operational costs, Anthropic is trying to lock down high-paying corporate clients by packaging their models into highly specialized scientific platforms. Anthropic just introduced Claude Science, a dedicated digital workbench tailored for laboratory researchers and academic scientists. The system runs on their existing model stack, but the value lies in the specialized tools plugged onto the interface. Claude Science arrives pre-connected to dozens of major scientific databases, allowing the software to render detailed protein folds and chemical diagrams directly inside the browser window. The workflow centers on literature reviews, data analysis, and running intensive laboratory experiments in a single application. This marks the fourth corporate vertical Anthropic has built out over the past few months following specialized drops for legal, financial services, and small business operations. The corporate playbook is clear. Anthropic noted that Claude Science had to pass their strict internal biosecurity evaluation before release. Software companies do not typically run biological weapons checks on a standard text tool. It proves the model is powerful enough to extract dangerous insights when connected to raw research archives. Management feels the risk surface is real and they are stating it publicly to win the trust of regulated enterprise buyers. Anthropic is using software verticalization to win enterprise trust, but Amazon Web Services believes the only way to lock down corporate cloud deals is to send physical engineering teams straight into your office. Amazon Web Services announced a new corporate division called the Ford Deployed AI Engineering Team, backed by an initial $1 billion capital allocation. AWS plans to send small engineering pods to work directly on-site with corporate clients. These cloud engineers build custom intelligence pipelines alongside the client's internal staff, train the corporate team on the new systems, hand over the operational controls, and exit. Initial enterprise clients include the NBA and Rico, and Amazon expects to scale the division to thousands of engineers. To understand why Amazon is spending a billion dollars on on-site consulting, look at the recent enterprise data. Over half of major corporations now have AI running in production, but jumping from a simple API key to a system that manages real corporate logic is incredibly painful. It requires hooking the model into ancient legacy databases and navigating resistance from internal teams. Amazon is turning this operational gap into a premium service. The strategy contains a brilliant competitive lock-in. The deeper these AWS engineers embed themselves into your daily corporate operations, the harder it becomes to switch cloud providers later. Human integration is a much stronger corporate moat than chasing minor variations in model performance benchmarks. While Amazon sends engineering pods to secure corporate server deals, a startup in Singapore is trying to bypass separate applications entirely by embedding an AI agent directly into your smartphone keyboard. A Singapore-based startup named Acti just launched an AI agent that lives directly inside your phone's native keyboard app. While you are mid-conversation in a messaging thread, you can hit a dedicated key to let the agent run tasks across your other mobile applications without ever forcing you to minimize the window. ACTI calls these automated macros skills. The system runs on Google's Gemini models and operates local first, keeping your private messages on the device unless you activate a feature that requires a cloud server. CEO Young Wang knows this distribution model well, having previously scaled the FaceMoji keyboard app inside Baidu. The company just closed a $5.3 million seed funding round led by Bitcraft Ventures. A small seed round deserves attention alongside billion dollar chip deals because the keyboard is the primary gateway to mobile software. You use it constantly across every application on your device. If automated agents are going to capture everyday consumers who do not care about technology, a hidden tool inside the keyboard is a much better entry point than a standalone chat app. Chat interfaces force you to remember they exist and manually open them. The keyboard is already active under your thumbs. Acti has a long road to prove its scale, but the interface choice is incredibly sharp. And that's it. If you have any feedback about this show, you can email Mike at yesterday andai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. Thanks for listening to this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.