AI News: 1X Neo, Extropic’s TSU, MiniMax M2 and Cursor 2

1X has opened preorders for Neo, a humanoid robot intended for the home. Its launch video went viral, and I have already preordered one. In October 2025, the company was offering it for $20,000 with availability planned for early 2026. I hope to test whether it can help with laundry, dishes and tidying when mine arrives.

The specifications shown in the launch material list a 66-pound robot, a 154-pound lift capacity, 22 degrees of freedom in its hands and a maximum noise level of 22 dB. These are the company’s specifications, not my own test results. Some initial reviews of the demonstrations say that many tasks still rely on teleoperation; I wonder how many people and how much infrastructure will be needed to provide remote assistance.

Extropic has released more information about its thermodynamic sampling unit, or TSU. Its pitch is that today’s computing uses too much energy, and that new probabilistic hardware and software can improve efficiency. Rather than executing conventional deterministic instructions, the proposed system samples from probability distributions.

Extropic claims the TSU could be up to 10,000 times more efficient than traditional CPUs and GPUs. The work is early: these are prototypes and simulations, not a demonstrated universal advantage in production. I am excited to see a novel computing approach tested, even though the details are difficult to understand.

Artificial Analysis reports that MiniMax M2 reached a new high for an open-weights model on its intelligence index at the time. It has 200 billion total parameters, with 10 billion active. The displayed 2025 snapshot assigns it a score of 61, below Claude Sonnet 4.5 and above Grok 4 Fast and Gemini 2.5 Pro. This is a particular benchmark snapshot, not a universal ranking.

Nvidia took a $1 billion stake in Nokia, according to the CNBC report I discuss, and the companies announced a strategic partnership on 5G and 6G technology using Nvidia chips. I find the investment notable alongside other relationships between AI companies and chip suppliers; my speculation about those financial ties is not a finding of wrongdoing.

IBM has released Granite 4.0 Nano for edge and on-device use. The family includes hybrid state-space models of about 1.5 billion and 350 million parameters, along with traditional transformer versions in those size classes. IBM’s displayed chart compares general benchmark performance and model size; I find the small models promising for their size.

Elon Musk has floated using idle Tesla vehicles as a distributed inference fleet. His hypothetical calculation imagines 100 million cars with a kilowatt of inference capability each, yielding 100 gigawatts in aggregate. I find the idea of owners potentially earning money from unused compute interesting, but this is a proposal and calculation, not an existing fleet or a demonstrated payment scheme.

Cursor 2.0 introduces Composer, a model the company says is four times faster than similarly capable models and designed for low-latency agentic coding. Speed matters to me as much as quality because it affects how quickly I can iterate. This is Cursor’s comparison, rather than an independently verified result in this video.

The updated interface centers on agents rather than files. Cursor also describes code review and testing as emerging bottlenecks, and says its new tools make changes easier to inspect and include a native browser for testing and iteration. I see the design as another step toward building software through collaboration with an agent.

US startup Substrate says it is developing advanced X-ray lithography for a new semiconductor foundry. It shows patterns that it says match features printed at the 2-nanometer node and below. Those images are a company demonstration of features, not proof that the startup is already manufacturing complete chips at scale.

Microsoft promoted Copilot mode in Edge for Windows and Mac. At the time, OpenAI’s ChatGPT Atlas browser was available only on Mac. I speculate that the platform comparison could be a competitive signal, though it may not be intended as one.

Amazon announced plans to cut 14,000 corporate roles, with reporting linking the reductions to preparation for wider AI adoption. I am skeptical that AI alone explains the cuts and suspect previous overhiring may also be involved; that is my interpretation, not a verified account of Amazon’s decision. The company said it would continue hiring in strategic areas while seeking further efficiencies.

The report says most affected employees would have 90 days to seek internal roles, with severance and benefits for those who did not find them. I feel for people facing that transition. CEO Andy Jassy had said in June that AI would eventually mean fewer people in some jobs and more in other kinds of work; that does not tell us how many new jobs will replace the positions lost.

In a clip from my interview, AWS CEO Matt Garman encourages students to develop critical thinking, creativity and a willingness to learn. He argues that specific technologies can change over a career, while engineering education can teach people how to break down and solve problems. Those skills, in his view, help people adapt as technology advances.