AI News: Macrohard, Comet Plus and the Debate Over AI Welfare

Elon Musk says xAI is building Macrohard, a tongue-in-cheek name for a purely AI software company. I see the announcement as a possible step toward an operating system generated by an end-to-end neural network. That is my interpretation and prediction: the recruitment announcement itself does not establish that a working Windows competitor exists.

Musk suggests that a software company could in principle be simulated by AI. I think the hardware question is less important than whether a normally deterministic operating system could be generated dynamically without sacrificing consistency or introducing hallucinations. Recent interactive world demonstrations make the idea intriguing, but they do not prove that an AI operating system is practical.

Meta announced a partnership to license Midjourney's aesthetic technology for future models and products. Chief AI officer Alexandr Wang described collaboration between the companies' research teams. I speculate that Meta wanted outside help with image generation and hope users eventually get access to Midjourney features, but the announcement did not spell out a free product or specific release.

Dynamics Lab presented Mirage 2, a real-time generative world engine. Its examples turn text or an image into a playable scene, including a stylized village and an explorable version of Van Gogh's Starry Night. Users can share a created world by link; I cannot tell from this announcement that they can play together simultaneously.

The demos show navigation and actions such as running, jumping and attacking. A company comparison claims more than ten minutes of interaction and operation on a single consumer GPU, alongside a comparison with Genie 3. These are developer-supplied claims about particular demonstrations, not independently verified benchmark results.

Figure showed a robot folding towels from a simple instruction. I find it notable because manipulating flexible fabric is difficult: the company says the demonstration was autonomous rather than teleoperated. One selected demo does not establish that an affordable general-purpose home robot is ready, even though I am excited by the progress.

A political group called Leading the Future announced a bipartisan effort to support pro-AI candidates and policies, saying more than $100 million had been committed initially. Its supporters include Andreessen Horowitz, Greg and Anna Brockman, Ron Conway, Joe Lonsdale and Perplexity. I tend to favor less AI regulation, though what the group will actually accomplish remains to be seen.

Perplexity announced Comet Plus to give subscribers and their AI assistants access to participating publishers' premium content. The proposed standalone price was $5, with Pro and Max subscribers included. Perplexity said it would distribute subscription revenue to publishers after retaining a portion for compute, using visits, citations and agent actions in its allocation.

At the time of the announcement, Perplexity had not named the initial publishers, and I would have liked examples to make the offering easier to assess. The question is important: if an AI assistant reads and summarizes an article for a user, publishers may receive less direct traffic, so we need ways to compensate their work.

A Pokémon Crystal experiment reported that GPT-5 defeated Red in 9,517 steps, compared with 27,040 for o3 in the setup shown. The chart tracks badge progression rather than a universal measure of intelligence. I see task efficiency as important because fewer actions can reduce latency and cost and leave more budget for difficult reasoning; the outcome is specific to this game and test harness.

The Information reported a deal under which Meta would spend about $10 billion on Google Cloud infrastructure and related services. I interpret it as a possible bridge while Meta expands its own data centers, but that is my speculation, not a disclosed explanation of the agreement. The reported scale also reflects a commercial arrangement, not a guarantee of future AI performance.

Intel announced an agreement under which the U.S. government would acquire a 9.9% stake through an $8.9 billion investment. Intel said the funds would come from $5.7 billion in previously awarded but unpaid CHIPS Act grants and $3.2 billion from the Secure Enclave program; the planned stake was passive and did not include a board seat.

I am conflicted: public ownership of a private company raises questions for me, while domestic chip manufacturing has strategic importance and Intel has faced intense competition. My remark that Intel was the last company manufacturing chips in the United States was too broad; other companies also manufacture semiconductors in the country. The agreement's outcome for Intel was still uncertain at the time.

Microsoft AI CEO Mustafa Suleyman argued that focusing on AI welfare prematurely could encourage people to anthropomorphize systems, potentially worsening unhealthy attachments and other human harms. This is his position in a contested debate, not proof that discussing AI welfare causes psychiatric illness.

Anthropic has taken a different approach by studying possible model welfare under uncertainty. It said Claude Opus 4 and 4.1 could end a narrow set of persistently harmful or abusive conversations after other attempts to redirect failed. Anthropic explicitly said it remained highly uncertain about the moral status of current or future models; this feature does not establish that Claude is conscious or human.

Mistral highlighted strong results for Medium 3.1 on a chatbot ranking, including second place in the category shown and eighth overall at that moment. I find it impressive for a relatively small model, but leaderboard positions reflect specific categories, users and evaluation periods and should not be treated as a permanent overall ranking.

Uber CEO Dara Khosrowshahi argued that driverless cars should be several times safer than humans and that, in the near term, reaching that standard with cameras alone would be difficult. He favored combining cameras, radar and lidar, whose cost he said had fallen. He allowed that a camera-only solution might eventually be possible.

Musk responded that lidar and radar could reduce safety through conflicting sensor inputs and claimed Waymo could not drive on highways. I am skeptical of that explanation and think city driving may be harder than freeway driving, although that is my intuition. The companies' competing claims do not by themselves settle which sensor architecture is safest; deployment area and time also matter.

NVIDIA announced Jetson Thor for robotics. The company lists up to 2,070 sparse FP4 teraflops, 128 GB of memory and a configurable 40–130 watt power range, and claims up to 7.5 times the AI compute and 3.5 times the energy efficiency of the preceding Jetson Orin generation. These are NVIDIA's specified comparisons, not expected speedups for every robot workload.

I discuss a preliminary Project NANDA report often summarized as 95% of generative-AI pilots failing. The report itself describes limited measurable business returns and a gap in deployment, which is narrower than proving that 95% of all pilots failed or produced no value. Its methodology notes a review of over 300 public initiatives, interviews with 52 organizations and survey responses from 153 senior leaders; the numbers in the news account I read differ.

My hypothesis is that some companies rush into AI because of hype without identifying a useful problem, understanding the technology's limits or testing how it fits into real workflows. That is an interpretation, not a causal finding proven by the headline statistic. Failed experiments can still teach teams which applications are worth pursuing.

I remain optimistic because I have seen useful applications in my own life and business. But assessing whether a company benefits requires a clear use case and an appropriate measure of impact, not merely a pilot launch or a sweeping percentage.