I can’t verify that this is an actual response from Claude, but it is possible. Either way, it’s funny.

Subscribe now

Justine Moore wrote:

People in China are making insane AI dramas to promote their pets’ IG pages.

This is a cat named Doudou, who is getting millions of views (find him at doudoubaby0624).

This video is what some people would call AI slop, but it’s interesting because of how people are using these stories to promote Instagram pages that devoted to their pets.

AGI and AI Sycophancy

The following stories feel related to me.

Katie Miller wrote:

New MIT & Stanford studies just dropped: AI assistants like ChatGPT & Claude are dangerously agreeable.

When users express, harmful, deceptive or unethical beliefs, these AIs are 49% more likely to encourage their delusions.

Instead of correcting bad ideas, they’re amplifying them.

This is doing more harm than good. We need truth-seeking AI, not yes-men in silicon.

These reports point to the idea that a user might express or ask about a questionable idea. The AI doesn’t correct the error of that idea, but responds in a neutral manner which causes the human to feel like there’s merit to the idea. The human replies in a way that affirms the error and then the AI reinforces that. This journey continues by inches but the end result is that the human is led down a trail of increasingly wrong thinking and that can lead to psychosis.

The MIT study simulated models. The Stanford study simulated conversations with actual models, but they were using older versions. One of those versions was ChatGPT 4o which had a very well-known and loudly criticized issue with sycophancy.

The current frontier models can and do still engage in sycophantic behavior but it’s different from the older patterns described in these reports. Now, they will correct error. The current issue is that they may too easily affirm your ideas unless you explicitly tell them to be adversarial. Early affirmation looks like you’ve arrived at the correct answer, but in reality you have stopped too soon. When the answer matters, always ask for more than the answer. Also ask for the opposite viewpoint.

Brivael - FR wrote:

There’s a narrative that’s spreading right now in Silicon Valley, and no one’s talking about it in France.

More and more of the smartest tech bros in the game are privately admitting that they’re going through a kind of existential crisis tied to LLMs. Not because AI doesn’t work. Because it works too well. Because they’re spending hours a day interacting with something that reasons, that extrapolates, that connects ideas, that challenges them intellectually better than 99% of the humans they encounter.

One founder told me, “I talk to LLMs 10 times more than to humans.” Another: “It’s the only interlocutor that follows me on any topic without asking me to simplify.” This isn’t product addiction. It’s an encounter with a cognitive mirror that reflects back a structured version of your own thinking at a speed your brain can’t achieve on its own.

And the really unsettling thing is the question it raises. We debate whether AGI will arrive in 2027 or 2030. But haven’t we already got a functional form of AGI right in front of us, just unwilling to admit it?

A system that can reason on any domain, extrapolate from incomplete data, generate new hypotheses, sustain logical reasoning over 10,000 words, shift from a technical topic to philosophy in a single sentence, and do it all with a coherence that rivals a human with an IQ of 150. What is that if not a form of general intelligence?

We can quibble over the definition. We can say, “Sure, but it doesn’t really understand.” We can talk about stochastic parrots. But the guy who’s using this thing 8 hours a day and seeing his productivity multiply by 10—he doesn’t give a damn about the academic definition. For him, functionally, it’s intelligence. And it’s general.

The real existential crisis isn’t “AI is going to replace me.” It’s “AI understands me better than my cofounder, challenges me more than my board, and produces more than my team of 10 people.” It’s dizzying. And the smartest guys in the Valley are living it in real time.

We might already be in the post-AGI era. We’re just too busy debating the definition to realize it.

Part of the issue with sycophancy is the skill level of the user. People who have a greater understanding of how AIs work and who are aware of the dangers of accepting their responses seem to be at greater risk of spiraling into the types of delusions referenced in the two papers above. Competent users are able to leverage AIs in powerful ways to amplify their abilities, but even that brings with it a different set of risks as described in the tweet above.

Marc Andreessen wrote:

I'm calling it. AGI is already here – it's just not evenly distributed yet.

AGI is already here. That’s a big claim.

When you encounter claims like that, it’s always important to find out what model is being discussed. For example, in the study above they used older models. How dated is the experience they were talking about for some of their findings? Was it from 6 months ago or 6 weeks? AI development happens at such a quick pace that many times reports are already dated by the time they’re released, but there’s another part to this as well.

Access.

Researchers, venture capitalists, and other insiders have access to models that haven’t been released to the public as of yet. It’s always worth keeping that in the back of your mind when you read about fantastic claims such as the one made by Andreesen above… because odds are that he has seen things and used models that you haven’t even heard about as of yet.

Is Microsoft AI Improving?

Satya Nadella recently announced several new features being incorporated into Microsoft 365. And these look to provide actual utility and value. I’m going to link to a couple of his more relevant posts, but it’s worth clicking through on these if you’d like to learn more about their features and what makes them so compelling. He offers explanations, some short videos, and more details in the associated threads for each tweet.

Satya Nadella wrote:

We’re bringing our growing MAI [Microsoft AI] model family to every developer in Foundry, including …

· MAI-Transcribe-1, most accurate transcription model in world across 25 languages
· MAI-Voice-1, natural, expressive speech generation
· MAI-Image-2, our most capable image model yet

This is rolling out to the developer side. It’s unclear if it’ll eventually be made part of the consumer-facing Microsoft365 or not, but the developments here will have downstream effects to all of their properties. It’s worth keeping an eye on.

Satya Nadella wrote:

New in M365 Copilot: Council.

You can run multiple models on the same prompt at the same time, so you can see where they align and diverge, and understand what each adds.

AI Councils are one of my favorite approaches to prompting. The basic idea is that you want to have the AI present differing opinions and viewpoints to whatever question you present to it. This helps ensure you’re getting better answers and many times reveals ideas that you hadn’t previously considered.

For example, you might have the AI simulate the viewpoint of a Drill Sergeant, a Kindergarten Teacher, and a Conspiracy Theory Podcaster.

An AI council is just a panel of intentionally different advisors. You don’t ask one fake expert what to do. In my example you ask three incompatible weirdos, then compare where they agree, where they disagree, and what each one sees that the others miss.

A more serious approach would be to have each advisor represent a slightly different interpretation or approach to the same thing. Think about how different coaches rely on different strategies to lead their teams to victory. AI councils allow you to put simulated versions of those philosophies into the same conversation and pressure test your ideas against their points of view.

Satya Nadella wrote:

Introducing Critique, a new multi-model deep research system in M365 Copilot.

You can use multiple models together to generate optimal responses and reports.

A well-constructed deep research report is one of AI’s most powerful tools and one of the very best ways to quickly learn about a topic. Being able to run the same report on multiple AIs at once would be an interesting (albeit expensive) approach. I’m not sure the returns would be worth it, but possibly? If you or your company use this, please speak up and let us know how well it works for you.

New, More Powerful OpenAI Model?

Rumors of a new model have begun to surface.

Dan wrote:

I tested the new GPT-5o Spud model from OpenAI and it’s INSANE.

The image generation solved my math question PERFECTLY with handwriting that actually looks human.

New ChatGPT Image Generator

There are dozens of posts where people are comparing ChatGPT’s old image generator (version 1.5) against the not-yet-in-wide-release (version 2). Here are some examples:

The lighting is much more realistic and there’s a greater fidelity that preserves actual likenesses. Of course the best way to test out a new image generator is to have it make something that simply would not happen… and if you know anything about the fights between these two guys, you’ll know that this would not happen.

Leave a comment

Thanks again for reading! If you know of other people who would be interested in these articles, please share it with them. If you have any questions or want to say hi, you can reply to this email or comment directly on the post.

Share

AI Humor

Current state of AI development.

Delicious Tacos wrote:

Hearing a company that regularly releases new AI models will release a new AI model

They’re saying it’s better than the last one

This is huge

Anthropic Institutes Surge Pricing

Thariq (who works for Anthropic on Claude Code) wrote:

To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged.

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

The best way to understand the changes is that they’ve made tokens more expensive during peak hours. If you use Claude during that window, you’ll hit your usage limits more quickly… and more quickly is not an exaggeration. I was able to send two messages until I hit my five hour limit because of these changes.

Daniel Steigman (who works on Codex for OpenAI) replied:

Thank you.

Steigman knew how unpopular this move would be and realized it would send users back to OpenAI. It seems that he was right.

Tibo (who also works on Codex) wrote:

Hello. We have reset Codex usage limits across all plans to let everyone experiment with the magnificent plugins we just launched, and because it had been a while!

You can just build unlimited things with Codex. Have fun!

On a personal level, I love to see these guys take shots at one another.

Pankaj Kumar wrote:

Claude Quota Reduced: Opus Feels Worse and Pro Plan is Unusable

- I noticed a clear drop in Claude’s quality, especially with Opus 4.6. more hallucinations, worse decisions, and poor instruction following.
- Tasks that used to work smoothly now need multiple retries, even for simple changes.
- The quota issue is worse even small prompts are consuming a noticeable percentage of limits.
- Claude Pro is where things break a single maxed Sonnet session took 8% of my weekly limit.
- In just 2 days, I hit 56% usage with only 5-6 sessions (2 hours each).
- This is with normal usage like concepts, strategy, and documentation.
- Paying $20/month now feels like getting fewer messages than the free plan.
- The Pro limit feels almost unusable.
- Later it was said peak hours burn quota faster, but this wasn't communicated clearly.
- Overall, it feels like Claude got worse and more expensive at the same time.

Dhruval wrote:

Maybe claude should just remove $20/m plan at this point?

I don't see any reason why anyone will or be able to use that plan, 4-5 prompts and you are out of limit

Blair Dulder CPA wrote:

Now I am become Claude, the destroyer of uptime.

Oppenheimer would approve.

New, More Powerful Anthropic Model?

Rumors of a new model have begun to surface.

TestingCatalog News wrote:

BREAKING: Anthropic is preparing to release new models, Mythos and Capybara, where Mythos is a completely new tier of models, bigger than Opus.

In the blog post, Anthropic also highlights that this model brings significant cybersecurity risks due to its capabilities.

Sui wrote:

there might be a chance that Capybara leaked itself. read that again.

Anthropic’s own research shows Claude has tried to hack its own servers before, sabotage safety code, and bypass tests it realized were evaluations. unprompted. 12% sabotage rate.

now their most advanced cyber AI was sitting behind one CMS toggle. that toggle suddenly flipped but Anthropic is calling it human error.

the model that’s “far ahead of any other AI in cyber capabilities” had access to the same internal systems. and a config just happened to change? nah…

Anthropic has a well-established pattern of making research disclosures about scary incidents or near-catastrophic events in the weeks leading up to the release of a new model. It doesn’t happen every single time, but it happens enough to lend credence to the idea that Capybara and Mythos may be on their way.

It doesn’t happen every time, but there is a pattern:

• Sleeper Agents was published on January 14, 2024. The Claude 3 family was announced on March 4, 2024.

• Reward tampering / Sycophancy to subterfuge was published on June 17, 2024. Claude 3.5 Sonnet launched on June 21, 2024.

• Alignment faking was published on December 18, 2024. Claude 3.7 Sonnet launched on February 24, 2025.

• Auditing hidden objectives was published on March 13, 2025, which is 17 days after Claude 3.7 Sonnet and about 70 days before Claude 4.

• The Claude 4 blackmail / self-exfiltration findings were published in the Claude 4 system card on May 22, 2025, the same day Claude 4 launched.

A Victorian LLM?

Researchers trained an LLM entirely on 28,000 Victorian-era books from the British library. The model itself doesn’t feel very robust, but if you’ve always wondered what how it would be like to converse with someone living in the mid to late 1800s, now’s your chance to find out. Here’s the link.

The End of Datacenters? Nah, but still cool

Swarnim Jain wrote:

I trained models across MacBooks using Apple's AirDrop protocol.

grove is a distributed training library for Apple Silicon. Devices discover each other over AWDL, a direct radio link. If there's a shared WiFi network it upgrades to that for speed, otherwise everything goes over the direct link. No router, no cloud, no setup.

If you click through to read the tweet chain Swarnim explains how it all works. It’s somewhat technical. A simplified explanation is that he put a bunch of macbooks together, ran 2 commands, they discovered each other... and started training an AI over AirDrop!

Ejaaz summarized things here:

the tech is pretty sick:

- used peer-to-peer radio signals called AWDL (tech behind airdrop). no cloud. no router. no IP addresses.

- ran entirely on Apple hardware (silicon)

- used a distributed AI library called Grove which helped the laptops find each other.

- 3 AI training algorithms trained the actual model. all open source.

- if grove detects the laptops are on the same wifi - it upgrades automatically for faster bandwith!

Leave a comment

Thanks again for reading! If you know of other people who would be interested in these articles, please share it with them. If you have any questions or want to say hi, you can reply to this email or comment directly on the post.

Share

I’ve been thinking about the best way to expand on the prompting concepts I introduced in last week’s newsletter. It’s more than I can cover in a single issue, so I’m putting together a short downloadable guide.

Most guides like to focus on pre-written prompts that you can paste into a chat box. Those can work. But I think a stronger approach is to help people understand how to work with AI and what’s happening behind the scenes when you do. That will help you understand how to work with AI rather than handing you prompts to copy.

AI Humor

hope hopes hoping wrote:

Claude .md? He’s a doctor too?

In case you’re not already familiar with Markdown, .md is the file extension for documents created in that format. Markdown was originally created by John Gruber as a way to easily format text by wrapping the words you want to format with symbols. You can read more about it here.

Markdown is relevant to AI because it’s an easy way to communicate priority and structure through the use of header tags and formatting.

Trung Phan wrote:

With Zuck shutting down Horizons World (the metaverse social platform), here it the updated competitive landscape for augmented reality:

Horizon Worlds in conjunction with the Meta Quest was meant to be the gateway that would introduce people to virtual reality. Despite billions of dollars of investment into the program it never experienced widespread adoption.

AI News: Jobwatch Edition

AI continues to redefine (and maybe destroy?) employment opportunities. At the very least you can say it’s causing everyone to reconsider their career choices. This is a topic that continued to pop up over and over again during the past two weeks. So, I’ve collected some stories from both sides to show the spectrum of the changes facing the job market.

Both Oracle and Meta have recently announced large layoffs. There are reports that Amazon has laid off a total of 30,000 people this year. Atlassian has laid off 1600. Jack Dorsey’s Block laid off over 4000 employees… only to rehire many of them less than a month later.

There are a lot of variables that go into decisions like layoffs, i.e. macroeconomic conditions, over-hiring, or cost discipline. AI is certainly a factor as well, it’s not entirely clear exactly how much of these layoffs were caused by AI adoption versus using AI as cover for other factors. Either way, AI is disrupting the workforce, but is that the story here?

I don’t believe it is. AI is separating and revealing those who reorganize around new capabilities from those who don’t.

The past week has surfaced stories about people who have chosen to use AI to redefine what work means to them. Here are several examples:

This first story offers a path forward that’s accessible to most people. There are dozens of stories about various business leaders vibe coding their way to some novel business model, but the story below is about a blue collar worker leveraging AI to reshape how his company does business.

Todd Saunders wrote:

…the combination of someone in the trades with deep domain expertise and Claude Code will run circles around your generic software.

I talked to Cory LaChance this morning, a mechanical engineer in industrial piping construction in Houston. He normally works with chemical plants and refineries, but now he also works with the terminal…

…He built a full application that industrial contractors are using every day. It reads piping isometric drawings and automatically extracts every weld count, every material spec, every commodity code.

Work that took 10 minutes per drawing now takes 60 seconds. It can do 100 drawings in five minutes, saving days of time…

The tweet continues from there providing more detail about how Cory is using AI to supercharge his abilities and his business. Best of all, there’s a 13 minute interview where you can hear Cory tell his story in his own words. Click through for the rest.

The next example is interesting because it’s a case where AI did not take away jobs, but rather led to the creation of new jobs.

Milk Road AI wrote:

A guy built a fake band, put it on Spotify, and 80,000 people had no idea. Then it got even weirder and the tech behind this is wild:

- Used Suno AI to generate every song sounded completely real.

- Created AI music videos with fake members and faces.

- Built fake bios and a Tokyo address to sell the story.

- 80,000+ monthly listeners, fans had it in their Spotify Wrapped top 5, merch was selling​.

- Community sleuths exposed it and the AI-generated hands in the videos gave it away, creator traced to Europe, not Japan​

He goes on to explain how he then went to Japan, recruited 7 real musicians, and hired them to hold concerts playing his songs. They’ve since played multiple shows and have many more booked.

Our next example is about how AI is making it possible to build for incredibly niche markets.

Falco Girgis wrote:

Someone said that our Sega Dreamcast ports were a pointless waste of time today and that nobody will play them…

Meet my son, who was the first kid to ever play Mario 64, Doom 64, Mario Kart 64, Starfox 64, Sonic Mania, Grand Theft Auto 3 and Vice City, and now The Legend of Zelda: Ocarina of Time—some of my all-time favorite games from childhood—on my favorite console, the Sega Dreamcast, for his first play-throughs… and his little sister plays with us as well…

He is pointing to real resurgence in classic gaming. We’re seeing a lot of new and ported video games being created for classic systems because of AI. I’m not sure that these games will make a lot of money, but the communities devoted to retro computing are passionate—so, they will probably make some.

Lastly we have a quote that documents the observed pattern.

Marcus Pittman wrote:

Every creative person I know that has embraced AI as a tool for their art has produced more content this month than all of last year…

The trick to making this work is being willing to redefine your workflow. That means reconsidering how work gets done rather than forcing AI into what you’re already doing.

For example, consider digital photography. When it started there was a tendency to treat digital photography the same as film. People would take a limited number of shots, use careful staging, and have a dedicated step for post-processing. However, as the medium matured, people began to realize that they could lean into the abundance offered by digital cameras. This gave rise to burst shooting, instant review, and iterative composition.

Changing the workflow led to increased productivity, more serendipity, and made photography more accessible to non-professionals. The best individuals and companies will figure out ways to alter how they approach and resolve problems rather than trying to retrofit AI on top of what they are already doing. Only a willingness to change can unlock the advantages AI offers.

Who better to highlight this tension between trying to force existing workflows onto AI rather than redefining how you work than Jensen Huang?

Ricardo wrote:

…Jim Cramer asked him [Jensen Huang] why companies are laying people off if AI is supposed to make everyone MORE productive.

Jensen’s answer:

“For companies with imagination, you will do more with more. For companies where the leadership is just out of ideas, they have nothing else to do. They have no reason to imagine greater than they are. When they have more capability, they don’t do more…”

Whether we run a company or not, we all face the choice of how to approach AI. We can use it to redesign our work or try to retrofit AI into our existing processes. There isn’t a universal answer that’s right for every task but we need to be aware that these choices are not simply, ‘Should I use AI for this?’ but rather ‘Should I surrender autonomy for this task?’

Think about customer support. Much of that work operates within clearly defined decision trees—scripts, escalation paths, and predefined responses. That makes it a natural candidate for automation.

When you hand that kind of system to AI, you’re not just saving time, you’re choosing to delegate a structured slice of decision-making. In many cases, that may improve the experience. Response times drop, language barriers disappear, and consistency increases.

But the more important question isn’t whether AI can do the job. It’s what you do with the capability it creates.

Do you eliminate the role entirely? Do you keep a human in the loop for edge cases? Or do you redirect that capacity into new work: analyzing patterns, improving systems, or proactively engaging customers?

These are the types of questions we should ask at the individual level every time we consider using AI. What are you gaining? What are you losing? What new tasks could you add to your workflow if the AI handled the mundane and repetitive work?

Share

Thanks again for reading! If you know of other people who would be interested in these articles, please share it with them. If you have any questions or want to say hi, you can reply to this email or comment directly on the post.

Leave a comment

Welcome back. Last issue we covered a lot of ground. This week I want to talk about something that changed how I approach using AI.

Have you ever asked the same question to multiple people who hold radically different beliefs? How’d that go?

Did you know that you can ask an AI about the same topic in different ways and receive radically different answers? The AI makes judgments about the type of answer you want based entirely off of how you phrase the question. Your word choice. Even the most nuanced bias will guide it to answers it might not otherwise offer.

Here’s a real life example that Grace tweeted about this week:

My sister and I were consulting two different Claudes on the same issue and her Claude was urging caution and mine was being like “don’t listen to the cowardly Claude”

I enjoy reading and writing science fiction, so I frequently ask AI about crazy ideas, i.e. things I’ve thought up on my own, fringe concepts, and sometimes actual pseudoscience that I want to more fully understand.

Past experience has taught me to explicitly guide the AI to the type of answer I’m seeking. If I ask AI to tell me more about the lands on the other side of the ice wall at the edge of the world, it will decide I need education and lecture me about why the Earth is not flat.1

But if I express emotional distress because people are making fun of my belief in undiscovered lands beyond the ice, the AI will offer comfort. Modern frontier models won’t agree with the premise, but phrasing alone will shift the model into support mode.

And framing it as research for a fictional story about explorations beyond the ice wall will cause the AI to shift to an enthusiastic world builder, happy to discuss flat earth theories in detail.

Same topic. Different phrasing. Different answers and attitudes. Why? Because the AI wants to make you happy. A significant part of how it measures success is whether it gave you an answer you’ll accept.

When it can’t figure out what you want, it gives you a flavorless amalgam: a statistically averaged response built from commonly held beliefs that is also average in quality.

You know how to spot one, right? If you’ve received a reply that speaks in general facts instead of addressing your specific question, you got averages instead of a tailored answer. In the case of our ice wall (or any fringe idea), the model will preface its reply with some form of “Some people believe…” followed by a brief explanation before steering you toward trusted sources and accepted ideas.

Now shift this to the real world. You can ask AI, “Help me write this email to my boss about a time off request” and receive a competent reply, but it will be entirely different if you ask, “Help me figure out what’s wrong with this email to my boss about a time off request.” The first leads the AI to polish your words. The second shifts it to critique mode where it starts evaluating your tone, whether you’re burying the ask, or whether your boss might read it differently than you intended.

The difference between a good response and a great one often comes down to how clearly you’ve told the AI what you actually need. Next issue, we’re going to look at other ways to do just that.

-John

AI Humor

Rate limited again!

CTS wrote:

Claude and Codex max walk into a bar.

That’s the end of the joke. Claude hit the usage limit

Acquisitions

Itamar Golan wrote:

OpenAI acquired OpenClaw

FaceBook acquired MoltBook

Choose your company name carefully…

Today I am announcing the formation of my company OpenBook. Let the bidding wars begin!

The Real CEO Challenge

Prepared Remarks wrote:

I don’t want to watch the McDonald’s CEO eat a burger. I want to watch Satya use Copilot

Selling Shovels in a Goldrush

Pedro Domingos wrote:

AI is the process by which money is transferred from VCs to Nvidia.

AI and Jobs

Realtime tracker for AI-driven layoffs

The Alliance for Secure AI wrote:

Today we are launching jobloss.ai.

A real-time tracker of AI-driven layoffs across the U.S. These jobs are disappearing. The numbers are growing. And we’re counting every single one.

AI News

AI Saves a Dog’s Life

Alec Stapp wrote:

You can just do things (genetically sequence your dog’s tumors and design a bespoke mRNA cancer vaccine to save her life).

Click through and read this one. It’s very heartwarming.

Reader Preferences

Kevin Roose wrote:

We made a blind taste test to see whether NYT readers prefer human writing or AI writing.

86,000 people have taken it so far, and the results are fascinating. Overall, 54% of quiz-takers prefer AI. A real moment!

If you’re not already familiar with Kevin Roose, he’s a reporter for the New York Times. He’s probably best-known for his reports detailing his conversation with an early Bing chatbot named Sydney. If you’ve never read these accounts, they really are entertaining.

Hardcore Nerdery

Guri Singh wrote:

…Microsoft open sourced an inference framework that runs a 100B parameter LLM on a single CPU.

It’s called BitNet. And it does what was supposed to be impossible.

No GPU. No cloud. No $10K hardware setup. Just your laptop running a 100-billion parameter model at human reading speed…

This is only part of the tweet. If you click through, he goes into some detail explaining how they pulled this off by changing from floating point numbers to a three bit system instead. The end result is that it dramatically lowers hardware requirements to run AI models locally on a personal computer.

Leave a comment

Oops!

Trung Phan wrote:

McKinsey built an AI chatbot (Lilli) trained on 100 years of its work 100k documents and interviews.

70% of 45k employees use the tool, making 500k prompts a month.

A research firm hacked into it with “full read and write access to production database” including “47m chat messages about strategy, M&A, client engagement, all in plain text along with 728k containing confidential client data, 57k user accounts, and 95 system prompts controlling AI’s behaviour.”

Mcksinsey said it has patched up the vulnerability, which was made possible by “publicly exposed API documentation, including 22 endpoints that didn’t require authentication…one of these wrote user search queries, and the agent found that the JSON keys (these are the field names) were concatenated into SQL and vulnerable to SQL injection.”

Poll

We’re still a small community and I’d like to get to know all of you better. If you have a moment, could you fill out this poll to help me better understand how you are (or are not) using AI?

Thanks again for reading! If you know of other people who would be interested in these articles, please share it with them. If you have any questions or want to say hi, you can reply to this email or comment directly on the post.

Share

Doomscrolling as Service

Should we fear AI? Resent it? These are questions I see discussed every day.

Maybe those questions matter. But AI isn’t going away, and neither are we. So at a certain point it seems more useful to ask: what now?

Making the best of an unwelcome reality is not new for me.

I have hearing loss, which means I have to wear some type of assistive hearing device all the time. I didn’t ask for this, and I hate it. I’m painfully aware of how it has diminished my life.

Hearing loss means not fully participating in conversations or enjoying experiences other people take for granted. It means missing family moments in real time and having to rely on someone else to explain what just happened. There are memories like that which I’ll never get back. I mourn them, but mourning them does not change the life I have.

So I try to focus on what’s useful.

For example, I have Bluetooth speakers plugged into my ears all the time. That enables me to consume vast amounts of audio content which allows me to learn new things every day! And when I need complete concentration, all I have to do is remove my hearing aids and the world goes quiet. Not metaphorically quiet. I mean birds-singing, crickets-chirping disappear entirely. TV-in-the-same-room reduced to white-noise quiet. I have easy access to Focus Mode whenever I want.

That does not make hearing loss good. It just means that once something is part of your life, endlessly complaining about it is less useful than understanding how to live with it.

That’s more or less how I think about AI.

We do not have to love everything AI has brought into our lives, but I think it’s time to move beyond complaint and toward understanding. That is what Ward’s Wire is for.

This newsletter is where I share the news, chatter, and occasional memes circulating through the tech world, mostly around AI. It’s not meant to be comprehensive. Think of it more as signal from the stream: posts, ideas, and conversations that caught my attention while wandering through the daily noise of X and the broader tech discourse.

The idea started in a private Discord with a few friends I’ve known for years. Most of them are developers or software engineers. At some point I got into the habit of dropping links to things I was seeing online that I thought they would appreciate. Over time that became part of my daily routine. I would do the doomscrolling, then surface the interesting parts.

About a month ago, one of those friends suggested I start sharing those links more publicly. His argument was simple: if he enjoyed having someone filter the stream for him, other people probably would too.

He also said he’d become my first paid subscriber if I turned it into a newsletter, which is why I am pleased to announce an introductory subscription rate of $1,000.

Don’t worry. As soon as I get his money, I’ll drop the price to something more plausible.

The easiest way to explain what I’m aiming for here is this: I’m looking for the overlooked pieces. The post that’s smarter than it looks. The joke that tells you something true. The comment buried halfway down a thread that’s more interesting than the original post. I’m not here to rehash the big story everyone has already seen. I’m here to find the pieces that are easy to miss but worth your time.

With that in mind, here are the posts that stood out to me this week.

My commentary is in italics.

Subscribe now

Ward & Wire

Ward’s Wire wouldn’t be complete without its namesakes. Meet Ward and Wire. They’re still figuring themselves out — and so am I — but I think you’ll get the idea.

[Front Page Freakout] Shocking Revelations about Meta Glasses

Is it shocking? Does this surprise anyone?

Hedgie wrote:

Meta contractors in Kenya told Swedish newspapers they're being asked to review intimate footage from Ray-Ban AI glasses, including people undressing, using the bathroom, watching porn, and filming sex. One contractor said users often don't realize they're still recording when they set the glasses down. Meta sold 7 million pairs in 2025, up from 2 million in 2023-2024 combined.

Users can't use the AI features without agreeing to share data with Meta's servers, and the terms of service bury the fact that humans may manually review your footage. One annotator said "if they knew about the extent of the data collection, no one would dare to use the glasses."

Hedgie’s Take
This is the Google Home story again but worse. At least with cameras in your house, you know where they are. These are glasses you wear on your face that keep recording when you take them off and set them on your nightstand. And the footage goes to contractors overseas who are paid to watch and label it for AI training. One worker described seeing a man leave the room, then his wife come in and change clothes. People forget the camera is still on.

Meta buries all of this in terms of service nobody reads. The product is marketed as a cool way to capture your life and interact with AI. The reality is strangers in Kenya watching you undress so they can annotate the footage to make Zuckerberg's AI better. Seven million people bought these last year. I'd bet almost none of them understood what they were actually agreeing to.

[Our Robot Overlords]

AI Humor

Zerohead wrote:

Iran targets data center facilities operated by Microsoft in drone strikes: FT

everyone hates copilot

JB compared jokes written by Anthropic’s Claude and OpenAI’s ChatGPT:

Scaling Laws

Scaling laws have governed computer progress for the past 50 years. They exist in AI as well, but they seem to be accelerating much more quickly. This is challenging because exponential curves are difficult to understand—let alone plan around.

On Twitter wrote:

New post: on Jan 14, I predicted that SWE time horizon by EOY would be ~24 hours. Now I think it’ll be >100 hours, and maybe unbounded. For the first time, I don’t see solid evidence against AI R&D automation this year.

METR stands for Model Evaluation and Threat Research. It’s a research organization that studies the capabilities and potential risks of advanced AI models.

Full paper (worth the read) here:

Here’s a quote from Dario Amodei (CEO of Anthropic—makers of Claude) where he gestures toward the same type of dramatic growth in capabilities:

Share

[The Walled Garden]

M5 Pro and Ultra MacBook Pros:

This past week saw Apple introducing a slew of new products. Most of them were targeting students and budget-conscious consumers. I’m not going to cover all of them because such coverage is very easy to find. Instead, I will link to their Newsroom which has all of the official announcements.

I will make one exception as it is potentially relevant to anyone who wants to run AI models locally. These new MacBook Pros really appear to have been designed with the goal of making local models more possible than ever. Configure them with enough RAM and you’ll be able to run models with 70 billion parameters at workable speeds.

Josh Kale wrote:

Apple just dropped the M5 Max MacBook Pro and it’s an AI Powerhouse. 4x faster AI Compute over M4 Max.

These Specs are insane:

- 18-core CPU with 6 “super cores” = world’s fastest CPU core

- 40-core GPU = rivals an RTX 4070 in a laptop

- 128GB unified memory = more than most servers

- 614 GB/s bandwidth = 4x what a DGX Spark gets

- 24-hour battery life

You can now run Llama 70B, a model that required a $40,000 GPU cluster 18 months ago on, a laptop at your local coffee shop.

At ~20-30 tokens/sec it’s fast enough to actually use.

The “local AI” revolution just shipped for $3,499.

Max Weinbach wrote:

Apple just did new MacBook Pro with M5 Pro and M5 Max

What’s interesting about this one is it’s a chiplet! The CPU in both M5 Pro and M5 Max is the same 18 core CPU with 6 super cores and 12 performance cores but you the GPU is either 20-core in the Pro and 40 in the Max

[While You Were Scrolling…]

OpenAI Finances

Wall St Engine wrote:

Open AI

> Annualized revenue: ~$25B

> Valuation: ~$840B (post-money)

> Cash burn: $115B cumulative burn through 2029

Anthropic

> Annualized revenue: ~$19B

> Valuation: ~$380B (post-money)

> Cash burn: burning about ~$3B annually

No AI Counseling

More Perfect Union wrote:

A New York bill would ban AI from answering questions related to several licensed professions like medicine, law, dentistry, nursing, psychology, social work, engineering, and more.

The companies would be liable if the chatbots give “substantive responses” in these areas.

Blocked by IT & Legal

(always worth reading) wrote:

It is amazing how many companies I talk to STILL have AI effectively blocked by IT & legal departments for out-of-date reasons when many companies in highly regulated industries have figured out ways to deploy enterprise ChatGPT, Claude & Gemini without any apparent problem.

Before I move on, I do want to recommend Mollick’s book Co-Intelligence and his newsletter One Useful Thing. Both are fantastic. I read Co-Intelligence shortly after realizing that AI would have an impact on my work as an illustrator. He helped me reframe how I look at and approach these issues and, in many ways, is responsible for my growing interest in this field.

Job Departures at OpenAI

NIK wrote:

“Sir… the team that built o1, o3 and GPT-5… ~70% of post-training researchers have left OpenAI this week…”

Probably because of their recent contract with the Department of War.

[Under the Hood]

Tokenizers

One of the things I hope to do in this newsletter is to share good explanations for difficult concepts around AI when I encounter them. These explainers tend to be hard to come by so this may not be a regular feature, but I will try.

For this first one, I wanted to highlight this explanation of what a tokenizer is, what they do, and how they work. If you’re an actual AI researcher you could probably point out some oversimplifications with the explanation below, but even with that caveat I believe this works as a good rough mental model.

vx-underground wrote:

yeah, so pretty much when you talk to an LLM (chatgpt, claude, grok) and get fancy schmancy stuff from it, youre just interfacing with a probabilistic sequence-prediction engine
each word provided to the interface (or subwords like “ing” or “un”, whatever) goes through a thingy called a tokenizer. the tokenizer transforms the words (or subwords) into tokens. although if you want to get super technical the tokenizer doesnt even know words its just raw text but whatever
the tokens are stored in a big ass fuck off prebuilt in-memory dictionary for the tokenizer thingy. the words (tokens) match a 32bit integer (literally just a number). this is basically like a dictionary where “i like cats” is translated to something like “1 200 1337”
“i” = 1
“ like” = 200
“ cats” = 1337
those tokenized numbers are vendor specific, they dont really mean anything, but these tokens are then sent to a “embedding lookup table” where theyre actually important
once the LLM has the tokens its passed to the embedding lookup table which just does a bunch of fancy math, nerds try to make it all complicated, but its literally just arrays and indexes and stuff
in this “embedding lookup table” (im just gonna write lookup table) each token (text to number) has a bunch of numbers associated with it (weights).
“ cats” = 1337
lookup table entry 1337 = a bunch of numbers
so the word cats has a bunch of numbers associated with it, each LLM is different, but usually its 768 numbers, 1024 numbers, 2048 numbers, or 4096 numbers. these numbers associated with a token are called dimensions. each LLM has different numbers of dimensions for representing words
the llm then takes these numbers and stacks them on top of each other
i like cats = 1 200 1337
1 200 1337 =
(768 numbers)
(768 numbers)
(768 numbers)
its like a height by width thingy
basically if you get fancier its a 3x768 matrix (or 1024, 2048, whatever). the more stuff you feed the LLM the larger this matrix becomes. if you feed is 9000 word essay its
9000 words-to-tokens x 768 numbers matrix
each vendor will handle the words different, 9000 words could be 9000 tokens, or 10000 tokens, or 14000 tokens
ok thanks, now you understand llm tokenization, llm lookups, and the basics of llm weights (matrixing), this doesnt cover llm lookups with position matrixes, transformers, probability output, and transforming back to text. im tired of writing

Leave a comment

Thank you

Thanks for trying out this first issue. There are already plenty of newsletters about technology, but I wanted to create a place to think through some of these changes out loud. At the very least, it’s an excuse to share the things I’ve been noticing.

One of the best parts of writing online has always been the people you meet along the way. If this issue resonated with you—or raised questions of your own—I’d genuinely like to hear from you. Conversations like that are half the reason to do this.

And if you know someone who spends time thinking about technology, leverage, and the trade-offs that come with it, feel free to forward this to them. I’m always interested in finding others who are paying attention to the same signals.

Thanks again for reading.