So you’d be forgiven for missing the news that Neuralink—Musk’s brain-computer interface startup—has successfully implanted computer chips in its eighth and ninth patients. The company announced the successful operations last Monday on X, and the patients have since come out as P8 and P9 on the social media platform.

Nick Wray is the eighth patient to get the Neuralink implant. The trial participant, who has ALS, tweeted a status report Sunday night in which he expressed gratitude and excitement for the experimental technology.

Posting as @Telepath_8, Wray says that from his first day with the implant, “I was afforded a glimpse, the slightest whiff, of the capabilities I will achieve with practice and it is so far beyond anything I could have dreamt of as an able bodied individual.” The technology has given him a “level of digital digital autonomy” he hasn’t had in years. Wray goes on to say that “if ALS is the price of admission to an opportunity of this magnitude, you pay it; gladly, willingly, and without hesitation.”

The opportunity Wray refers to is participation in the PRIME study, a clinical trial for Neuralink’s chip, called N1, and the surgical robot tasked with installing it. To be eligible for participation, you must have quadriplegia—limited function in all 4 limbs—due to either spinal cord injury or ALS.

As the company explains on its website, the N1 chip is part of a brain-computer interface that allows patients to control a computer solely with their thoughts—an ability the Neuralink team calls Telepathy. “There are neurons in the brain that carry information about everything we see, feel, touch, or think,” the site says. A special surgery robot places electrodes around the brain, near neurons to gather information about their activity. Following surgery, information is sent from these electrodes to the Neuralink app, which is able to translate the data into intents—actions that the patient is trying to take.

The technology can’t restore patients’ ability to walk or use their arms (yet). But it has given the trial participants a new lease on life. In 2016, “P1” Noland Arbaugh was paralyzed from the shoulders down in a swimming accident. He only had one year left in college.

Thanks to Neuralink, Arbaugh has been able to navigate the internet on his own and pick up learning again. He’s found work as a keynote speaker, allowing him to start providing for himself. Arbaugh tweeted that after his accident, “I never thought I would be able to help anyone ever again.” Neuralink’s technology has changed that.

In April, Bradford Smith—the third patient to receive the Neuralink implant—wrote in a post on X, “I am typing this with my brain. It is my primary communication.” Smith ends the post with a shoutout to the futuristic tech company’s founder. “Thank you @elonmusk!”

Apple researchers published a paper on Saturday that casts serious doubt on the ability of reasoning models like OpenAI’s o1/o3, DeepSeek-R1, Claude 3.7 Sonnet Thinking, and Gemini Thinking to actually reason.

The authors argue that due to issues of data contamination—models being exposed to test questions during training—and overemphasis on final answer accuracy, current methods for evaluating these models don’t accurately reflect their true reasoning capabilities. They propose a novel evaluation paradigm consisting of logic puzzles that would allow researchers to modify problem complexity in a quantifiable way.

They test two pairs of language models, DeepSeek-R1/R3 and Claude 3.7 Sonnet thinking/non-thinking. Each pair shares the same underlying base model, but the “thinking” models have undergone additional training to teach them to reason through problems.

The researchers find that the non-thinking models are able to match the performance of the thinking models on problems that are not too complex. For moderately complex tasks, the reasoning models begin to outperform their non-reasoning counterparts. And beyond a critical level of problem complexity, performance in both models completely collapses. Essentially, when a problem becomes too difficult, both reasoning and non-reasoning models are unable to generate a correct solution. Not only that, but models appear to give up when complexity reaches this critical point—even when they have more thought tokens at their disposal, they stop trying to find the right answer.

I plan to make a separate post that walks through the paper’s motivations, methods, experiments, and results in greater detail. But the main takeaway is to remain skeptical of big claims about capabilities of current AI models—especially when the speaker has a vested interest in the AI market.

OpenAI CEO Sam Altman: “We are past the event horizon”

In a Tuesday blog post titled “The Gentle Singularity,” Sam Altman claims “[w]e are past the event horizon”—the theoretical “point of no return” in AI development. The OpenAI CEO suggests that we’re now easing into the singularity, when AI surpasses human intelligence and begins to improve itself.

Altman’s post seems to be motivated, at least in part, by a desire to ease anxieties about what the world will look like post-AI. But in painting such a rosy picture, he risks coming off as condescending and out of touch. This is perhaps most apparent in Altman’s acknowledgement of white-collar labor automation: “There will be very hard parts like whole classes of jobs going away, but on the other hand the world will be getting so much richer so quickly that we’ll be able to seriously entertain new policy ideas we never could before.” It’s clear that Altman believes AI will be a boon to humanity; it’s less clear what this will actually look like.

Altman does address the question of cost and energy. He tells us that the average ChatGPT query uses about 0.34 watt-hours, or “what an oven would use in a little over one second.” He neglects to add that with over 122 million users logging on daily (and presumably for more than a single query), ChatGPT consumes about 39.98 million kilowatt-hours every day, enough to power the Empire State Building for 540 days.

Of course, nearly all modern activity involves consuming energy in some fashion. But when Altman proclaims that “intelligence and energy…are going to become wildly abundant” or that “[i]ntelligence too cheap to meter is well within grasp,” all while OpenAI builds huge new data centers with enormous energy demands, his words start to ring a little hollow.

Updates from Apple’s annual Developers Conference

Apple held its annual Worldwide Developers Conference this week, and presentations mostly centered around design updates. Liquid glass received a lot of airtime, while Apple Intelligence took the backseat.

So iPhone controls are now translucent and appear to reflect and refract light. In an interview with Tom’s Guide, Craig Federighi, Senior VP of Software Engineering, gives a very thorough explanation of the philosophy behind liquid glass. He describes a sense of harmony between the user interface and the physicality of the device, and says that throughout the design process Apple designers brought an array of actual pieces of glass into the studio, conducting materials studies to see how the glass interacted with its environment. It’s a rather beautiful moment in which Federighi’s passion for the tech really shines through, and it’s not a little reminiscent of the late Steve Jobs’ obsession with detail. 

I found myself thinking about Jobs a lot this week. Apple obviously has a fundamentally different approach from other tech giants when it comes to AI, and I wonder what the late CEO would think were he alive today. 

Rather than building a general-purpose chatbot, developers are integrating AI into the Apple ecosystem. This means that use cases are more narrow, but they’re also more practical. Detecting when a phone call is placed on hold, for example, and keeping your place in line until a human agent picks up on the other end. Or noticing there’s a major pileup on your daily commute and suggesting a detour that will still get you to work on time. This deliberate and targeted application of AI, with privacy placed at the forefront, seems prudent and reflects what I hope our AI-integrated future will look like. And though Apple certainly overpromised and underdelivered on intelligent Siri, I’d rather they release a reliable, accurate, and safe model a few months from now than rush to deploy before the technology is ready. Perhaps there’s a bit of projection here, but I’d like to think Jobs would agree.