I really don’t like the narrative that we have 18 months until AI leaves us in the dust. Derek Thompson, co-author of Abundance with Ezra Klein, has found himself in the same place. Instead of thinking we’re doomed by the machines, he thinks the real crisis will be people voluntarily degrading their own cognitive abilities by outsourcing their minds and letting their reading and writing skills devolve. In his recent post, The End of Thinking, he writes:

One might imagine, then, that the potential arrival of all-knowing AI would serve as a galvanizing threat, like a Sputnik moment for our collective capacity to think deeply. Instead, I fear we’re preparing for the alleged arrival of a super-brain by lobotomizing ourselves, slinking away into a state of incuriosity marked by less reading, less writing, and less thinking.

It’s a scary thought. Literacy is a cornerstone of how we summoned the modern world into existence. It lets us transmit ideas across space and time in ways far beyond what oral traditions could do. To let it atrophy now, in the age of AI, would be to fulfill Neil Postman’s warning in Amusing Ourselves to Death — handing over our best selves not to tyrants, but to distraction. Keeping it alive must be one of our highest priorities.

Jerry Neumann takes a deep look at the investments happening in the AI space and concludes that AI will not make you rich, at least if you look at it like previous innovations:

The way to invest in AI is to think through the implications of knowledge workers becoming more efficient, to imagine what markets this efficiency unlocks, and to invest in those. For decades, the way to make money was to bet on what the new thing was. Now, you have to bet on the opportunities it opens up.

Like the web in the late 90s, we’re definitely in the frenzy zone. There‘s almost certainly a bust of some sort coming. But the really interesting stuff is what will come next. The question I’m asking myself is: What can you do to be ready for that? That doesn’t just include learning how to use the LLMs well. It extends to looking for the second- and third-order opportunities that are just over the horizon.

Shlok Khemani has documented how the memory system in Claude works:

Claude’s memory system has two fundamental characteristics. First, it starts every conversation with a blank slate, without any preloaded user profiles or conversation history. Memory only activates when you explicitly invoke it. Second, Claude recalls by only referring to your raw conversation history. There are no AI-generated summaries or compressed profiles—just real-time searches through your actual past chats.

These searches are accomplished through the conversation_search and recent_chats tools. You can prompt Claude to use these by writing something like, “Can you tell me what we talked about in our last 4 conversations?” You can even get Claude to give you information about these tools by asking, “What tools do you have available for dealing with chat history?”

This is very different than how ChatGPT’s memory works using a combination of a set of user-controlled memories and a dense generated summary generated from conversation histories that aren’t user visible.

I’m not sure which approach I like better. I think I prefer how Claude gives more explicit control over what goes into the context window as compared to not knowing what is in the generated summaries of previous conversations that are updated on an opaque timeframe.

Sound designer and host of the Twenty Thousand Hertz podcast Dallas Taylor was at Apple Park last Tuesday for the big event and recorded an AirPods Pro 3 hand’s-on experience video.

The differences stood out immediately. Noise cancellation is far stronger, making it easier to hear details even in a loud room. The bass is tighter and the highs are brighter. The new foam ear tips feel secure and comfortable, while transparency mode is so natural it feels like nothing is in your ears.

It’s a ringing endorsement and sounds like a great update to the product. The AirPods have been a hands-down awesome product for Apple, and I’m really looking forward to getting my hands on this latest version.

Even though the latest frontier models have improved, hallucinations by large language models are still a common problem. These are the plausible but very false statements that models will reflexively generate when they can’t generate a true statement. OpenAI’s blog post about their latest research into hallucinations gives an approachable explanation for this:

Think about it like a multiple-choice test. If you do not know the answer but take a wild guess, you might get lucky and be right. Leaving it blank guarantees a zero. In the same way, when models are graded only on accuracy, the percentage of questions they get exactly right, they are encouraged to guess rather than say “I don’t know.”

As somebody who was pretty great at figuring out multiple-choice tests in school, going so far as regularly evaluating the distribution of answers on the sheets as well as what I knew about the test writers to guide my guesses, this explanation really resonates for me.

Beyond the analogy, the post and research paper cleanly illustrates how hallucinations are an outcome of the statistical mechanisms inherent in large language models, and not some weird and random glitches.

I’m a bit late to the JSON Feed party, but I finally got there and implemented it on my site. It was seeing the external_url property being nicely implemented in Reeder (my current feed reader of choice) that sold me.

I’d expect most people reading this to know what a feed reader is, but just in case: Matt Webb has a nice explainer.

One of my favorite Steve Jobs quotes is about how everything around you being made by people who were no smarter than you. At Farnam Street, Shane Parrish takes that concept one step further, saying that it takes irrational dedication to make things happen:

Every single thing we see, someone had to will it into existence against the entropy of the universe and the indifference of everyone else. That’s what the entire built world is… What truly separates people isn’t some magical talent, but an almost irrational commitment to pushing through pain that would break most people.

Status quo is easy. Real progress is never easy. And, most of the time, the hardest thing for me is just getting started.

One of the best things to come out of Visual Studio Code is the Language Server Protocol (LSP). It’s the magic glue that decoupled and standardized how language tools worked with editors and it’s what enabled VSCode to pick up support for so many languages so quickly after its release. Now, LSP is a core feature of all the major editors, and keeps the door open for new editors to come to market.

Following this same spirit, the team at Zed is introducing the Agent Client Protocol, and it promises to bring the same benefits to using AI agents in the development environment of your choice. Like LSP, it‘s based on JSON-RPC and supports streaming responses. In Zed’s blog post introducing ACP, Nathan Sobo writes:

Any agent that speaks ACP can plug into a powerful UI, where the user can follow the agent around the codebase as it works, control its access to tools and MCP servers, and review all of its changes in a multi-buffer with full syntax highlighting and language server support.

The first agent besides Zed’s own in-process agent to support ACP is Google’s Gemini CLI. I’m really rooting for this to take off and for tools like Claude Code (still my favorite agent) and OpenAI’s Codex to adopt this standard. The potential for a unified ecosystem of AI coding assistants that support all sorts of editing environments is exciting.

Kirby Ferguson, creator of the seminal Everything is a Remix series which explained creativity in the digital age, has dropped his latest video: Infinite Remix. It tackles the pressing question of our time by going back thirty years:

In the 90’s we began uploading ourselves, one kilobyte at a time. We’re emailing. We’re chatting. We’re blogging. There were a lot of cats involved. We want to find our people, see what they’re doing, show what we can do. We want to matter even if it’s just for a second… In just a few decades, we sent hundreds of billions of terabytes, almost the entirety of human expression upward into the cloud.

Kirby’s viewpoint, which he gives at the end of the video: AI changes the way we work, extending what we can do. He sees a future of collaboration between humans and machines. This resonates with how I currently see using AI to allow humans to do superhuman things.

I’ve long enjoyed Andrew Tanenbaum’s 1985 quip: “Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway.” It was top of mind five years ago when Dr. Ian McDonald and I helped Wayve get set up with Azure Data Box Disk as a solution for getting massive volumes of telemetry from their self-driving car testbeds into the cloud for machine learning model training. Our napkin math showed they needed 100x more bandwidth than they had available, even with a fiber link in their office. Shipping disks easily won, despite the latency.

My dear friend Bryan Jones recently lived this adage again when he moved 400TB of his science lab’s data across the country over two days in his Porsche Cayman. The napkin math on his trip — assuming 36 hours door-to-door, including 26 hours of drive time at legalish speeds and a bare minimum needed for rest and refueling both man and machine — puts his trip’s bandwidth at about 25 Gbps. If he could have fully utilized a 1 Gbps connection full-time, it would have taken 37 days to copy the data over the internet.

It’s amazing that, even in 2025, the largest of data migrations are still well served by four wheels and a lot of hard drives. Bonus points if it’s a Porsche.

Interested in how he is positioning Microsoft for AI-driven disruption, especially in light of Microsoft laying off around 15,000 people this year while posting record earnings, Om Malik took a deep look at Satya Nadella’s recent memo to Microsoft, which frames the layoffs as inevitable — and argues for embracing that pain early.

For years, the sector has been generous to its employees, offering unheard-of perks and placing a premium on skills such as software development. AI, however, inverts that relationship.

Satya’s memo outlines a zero-sum vision of AI disruption, one where companies have a license to increase profitability without so many of those messy human costs. I’m a fan of Satya and currently hold Microsoft stock, but I’m no fan of this approach to the near future. I may be naïve, but I remain convinced that there can be a positive-sum version of how this plays out.

Finally. Thirteen years after the RX1 was first released, and ten years after the RX1R II, Sony has released a video announcing an update to one of my favorite cameras. Packing the 61MP sensor, focus, and image processing from the A7R4 into a ultra compact body with the same lovely 35mm/f2 lens of its predecessors, it’s pretty much pitch perfect.

Now, some people are complaining about the price. At $5098 in the US, it’s pretty steep. But the RX1 was never cheap. The RX1R II was priced at $3298 ten years ago, which felt really steep. That’s $4333 in 2025 dollars with inflation. “OK,” you say, ”that’s cheaper, right!” Oh no, my friend. You’re forgetting about the new US import tarifs of 25% on the declared import value. You’d have to make several assumptions to calculate a true adjusted cost, but… it’s close.

So, no, the camera isn’t cheap. The value proposition won’t make sense for many. But, this was never meant to be a mass market camera and the price is a rational extrapolation factoring in inflation and tariffs. If it made sense to you 10 years ago, it makes sense now. If it didn’t then, it still doesn’t. 🤷

Anyway, I’m stoked about this new release of an old favorite.

Rands (aka, Michael Lopp) has been on his own deep dive into the world of the robots (aka, AI) and how they affect the future. There’s good news and bad news, he writes, and we’re in that inflection point where we can’t yet grasp the long-term implications of the tools that are suddenly at our disposal.

Like social media before it, we can’t conceive of the long-term implications of these tools suddenly being at our disposal. We’re either giddy with excitement about their potential or digging deep to find any and all reasons not to trust them. But trust your instincts.

Maybe the most reasonable thing right now is to be both giddy and skeptical as hell. Go all in, and yet trust very little. We thought we knew what the Internet would bring us in the mid-to-late 90s, and we weren’t even close to knowing it all. We got some things right about five or ten years too early, and yet made ludicrous guesstamates about what might be possible that now seem juvenile.

Brie Wolfson travelled to Kevin Kelly’s studio in Pacifica to have a chat about how to have impact in the world and be happy about it. Their conversation explores alternatives to the Silicon Valley mantras of maniacal focus and relentless pursuit, advocating instead for a career full of creativity rather than a linear climb up a ladder. In case you don’t know Kelly, he helped start WELL, one of the first online communities, in the mid-80s and was co-founder of WIRED in the 90s.

I asked Kelly about the tradeoffs of focusing on a single thing if you want to be great… “Greatness is overrated,” he said, and I perked up. “It’s a form of extremism, and it comes with extreme vices that I have no interest in. Steve Jobs was a jerk. Bob Dylan is a jerk.”

As someone who has certainly had an interesting path through life, bouncing around between software, writing books, and photography, I love this permission to work differently. To measure success by internal satisfaction rather than just net worth or employee count. There’s value in being a generalist who follows their interests, even if it means abandoning things when they no longer serve you.

If you were an avid reader of WIRED in the 90s — you know who you are — you’ve read Peter Leyden’s work. He’s now working on a series of talks and essays called “The Great Progression: 2025-2050.” His thesis is that the arrival of AI, the rise of clean energy technologies, and the emergence of biotech add up to one of the great transitions of history. A transition that’s on par with what we saw in the post-war era, the gilded age, or even the founding of the United States. And, interestingly, each of these transitions also came with the kinds of social upheaval we see today.

Building on an existing partnership, Anthropic worked with Rick Dakan and Joseph Feller to create a free 12-lesson course around Dakan’s and Feller’s AI Fluency Framework. The framework consists of four interconnected competencies: delegation, description, discernment, and diligence, and emphasizes human-AI collaboration over just understanding AI technology.

Even if you’ve been working with AI for a while, I think it’s worth spending the 3-4 hours the course takes to hone your thinking about how to work effectively with AI.

Martin Fowler thinks that working with LLMs represents as fundamental a change in software development as the shift from assembly code to higher-level languages like Ruby and Fortran was. But unlike those transitions, this isn’t just a jump up in abstraction level. Martin writes:

When I wrote a Fortran function, I could compile it a hundred times, and the result still manifested the exact same bugs. Large Language Models introduce a non-deterministic abstraction, so I can’t just store my prompts in git and know that I’ll get the same behavior each time. As my colleague Birgitta put it, we’re not just moving up the abstraction levels, we’re moving sideways into non-determinism at the same time.

I think that how much of an impact this has will depend on where you use LLMs in your software development process. When you’re working with an LLM to create code, but then committing the results to be used, the line of non-determinism stops well within your view. On the other hand, if you’re using prompts to write code and not looking at the output, then you’ll have to sort out what systems you need to provide accountability that the final result meets your requirements, even if you’re not able to inspect every line of code.