By now, everyone with even a remote interest in AI has heard about OpenAI’s rogue AI agent hacking into the systems of AI company Hugging Face. Turns out, Hugging Face wasn’t the only victim:

Hugging Face was thought to be the only victim of the unprecedented hack - but OpenAI now admits its bot attacked several “publicly-available services”. The out-of-control AI found four logins online which allowed it to access four separate, unnamed services.

Personally, I am much less interested in the specifics of the attack (in case you are, Hugging Face has published a long and detailed analysis of the attack on its blog) and rather in the question of how, specifically, the agent broke out of its containment. And, much more interesting even, what this means for threat actors running their agents on systems and infrastructure which don’t even have such a thing as a harness to begin with.

Honestly, if your company has servers and data exposed to the internet (and who doesn’t), sleeping soundly at night, knowing that your cyberdefenses will hold up, might be a thing of the past.

↗ Link

As educational institutions globally grapple with the question of how to prepare students for a future where AI will be a ubiquitous part of the workplace, the team at the Law School at the University of Chicago just published a thoughtful plan on the topic. Key element: strip devices from 1L courses to protect foundational skills while mandating AI use in later ones.

We will be piloting a coordinated approach to classroom and examination policies for the core 1L curriculum during the 2026–2027 academic year. Across all 1L sections, we will prohibit the use of electronic devices such as laptops, tablets, and phones in the classroom. There will be some limited exceptions to this policy.

All in all, definitely one of the better approaches to AI in higher education I have seen.

↗ Link

By now it shouldn’t come as a surprise that most AI projects (especially of the corporate ilk) are abject failures. Here is a wonderful, and rather complete, write-up by Australian consultant Nik Suresh. Highly recommended reading for anyone in the corporate world wondering what the heck is really going on with AI.

I strongly believe there are entire companies right now under heavy AI psychosis and it’s impossible to have rational conversations with them about it. […] The reality is thus: the people in charge either have no plan, or see no path forwards other than keeping their heads down. […] In fact, it is in the interests of almost every actor in the space – boards, executives, employees, vendors, consultants – to obfuscate and misrepresent the success rate of AI projects. Many publicly traded companies are putting out announcements about their AI productivity gains when I know for a fact that the businesses have done nothing other than purchase Copilot licenses and declare victory.

↗ Link

Recruiting (and more generally the whole hiring process) is one of the (few) business functions that has seen AI tools being deployed in actual production environments: from AI-powered hiring target research, recruiting outreach, and candidate screening to resume parsing and onboarding (and everything in between). It turns out AI’s (usually) hidden biases are problematic (no one is surprised) – but they also have the wonderful capability of reinforcing themselves.

Researchers already know that LLMs pick up human biases from their training data. New research suggests that LLMs can also develop their own biases from experience – and stereotype job applicants more than humans do.

↗ Link

Researchers in the UK managed to poison an open-weight AI model (open-weight models have their trained parameters publicly released and made free to download – typically being used to run on a firm’s internal infrastructure) for $100 and in about an hour.

“Even when model weights are public (‘open weight’), we have almost no ability to predict its behavior,” they wrote. “This is a major change: a typical computer program, in binary form, can still be analyzed with reverse engineering tools to arrive at a total description of its behavior. With models, we have nowhere close to this capability.”

The skinny here is that this leaves you in a peculiar bind: use a closed-weight model (where you have no insight into the model’s behavior – such as OpenAI’s or Anthropic’s models) and trust that the vendors do everything they can to keep their models safe, or run open-weight models on your own infrastructure but make yourself potentially vulnerable to supply chain attacks (very similar to what we have seen numerous times lately in open-source packages for popular programming languages such as JavaScript).

Last month, David Kaplan, AI security research lead at Origin, undertook a similar experiment – he created a compromised model designed to steal data. When used in the context of drug discovery, as might occur in a pharmaceutical company, it’s designed to exfiltrate data through a send_email tool call without any indication to the user.

↗ Link

Here is something to mull over: L. M. Sacasas makes the argument that AI reshapes your perception and judgment from the inside, even when you use it carefully or reject its every suggestion. This leaves you in a position where even carefulness offers no protection – the most deliberate users lose judgment fastest because the medium, not the content, does the reshaping.

But AI is not a tool in this sense, it is an environment which envelops the user and works on us from the inside out while we naively think that we remain unchanged by our use so long as we are using it carefully and intentionally. The care and intentionality is beside the point, and our confidence in such vigilance probably works against us in the long run. […] And this is because AI is an environment not a tool. I can pick up a tool and put it down, but the environment absorbs me into itself.

↗ Link

I do like my LLM. Actually, I love it, and it’s insanely useful to me and my work. I also believe that there is a lot of potential in the further development and deployment of AI. But SoftBank’s Masayoshi Son going on the record to say that the AI buildout will require $5 trillion per year by 2040 and that AI revenue will hit 20% of global GDP by 2040 is just a bit too much for me…

“Every year $5 trillion, or 800 trillion yen, you might think that’s a lie, but I am confident that’s what it will cost,” Son said ⁠at SoftBank’s annual corporate conference in Tokyo. […] “The business model will be viable because by 2040, if AI revenue makes up 20% of global GDP, spending 800 trillion yen a year is a rounding error,” Son said.

The important bit here is this:

He did not say how he came up with the $5 trillion number or the proportion of global GDP he expects AI will make up.

Yep. Precisely. And just to put this all in perspective: Global GDP today is ~$117 trillion. 20% of that is ~$23T – larger than China’s entire economy ($19.4T) and about three-quarters the size of the US ($30.62T). All of worldwide manufacturing is about 15% of global GDP (and the single largest sector today).

↗ Link

The world’s insatiable appetite for AI is driving up the cost of RAM (computer memory) – and not by a little bit, but by a lot. You might have seen that your beloved Apple MacBook Air just got a whole bunch more expensive. Now, you might shrug this off as a minor inconvenience – at the end of the day, how often do you really buy a new laptop? But here is where this really bites: The world runs on cheap (Android-based) smartphones. If you have ever been in a developing country, you will have, inevitably, seen even some of the poorest people with a smartphone. And for them, the smartphone is not a toy to keep them entertained; it’s an important tool in their lives. And exactly these cheap smartphones are being squeezed out of existence. And that’s a real problem…

After ravaging the PC market for both DIY builders and OEM buyers, wrecking the game console ecosystem, and savaging the enterprise server landscape, the global memory shortage is officially claiming its next victim: the budget smartphone. According to new market analysis from tech research firm Omdia, the smartphone market, particularly the entry-level market, is on the brink of a massive contraction. Omdia projects that global shipments of smartphones priced below $400 will plummet by over 22% this year, dragging the entire global smartphone market down by 12% year-over-year.

↗ Link

Scott Alexander, on the Astral Codex Ten blog, makes a strong argument for AI getting better and better at forecasting – and now regularly competing with (and sometimes outperforming) human superforecasters. The idea is simple and logical: Given that AIs have been trained on vast amounts of data, with the right harness, they should be able to make predictions with calculated confidence levels.

Ironically, we’ve spent the last couple of years making AI refuse to have opinions, and now we’re building a certification layer specifically so it can – which also means that the question isn’t whether AI forecasters are accurate; it’s whether we’ve thought at all about who decides which opinions get to count as “calibrated” versus “biased.”

But the AI prediction markets of the future should be far superior to the human markets of the present. The biggest barrier to current markets is liquidity - there isn’t enough money riding on most questions to convince top superforecasters to drop their jobs at Google or the NSA in order to think hard about them and bet on them. But AI forecasters bring the cost of forecasting labor down to near-zero, so we can have hundreds of different AI agents betting on each question and be pretty sure its error has been driven down to the theoretical minimum. This, in turn, means we can vastly expand the number of questions, including (finally!) allowing randos to submit their own questions (probably with AI assistance in proposing un-rules-lawyerable resolution criteria). […] This, then, is my prediction for the AI superforecaster future: for basic questions, your off-the-shelf AI chatbot will be able to offer opinionated probabilities superior to those of any human. For more controversial or bias-laden questions, a new era of prediction markets will smooth over differences in brand and model and efficiently aggregate all AIs’ opinions.

↗ Link

Silicon Valley legend Kevin Kelly’s latest think piece is about our relationship with omnipresent AI assistants. In Kelly’s view – and remember, Kelly was a pioneer of the quantified self movement – we will soon all be running around with AI companions, which will become an extension of ourselves – hence the term “exoself.” In Kelly’s view, we will be making choices between four types of exoself relationships: Twin/Clone, Tutor/Guardian, Counselor/Assistant, and Hero/Friend.

So far, so dystopian (good or bad – you decide). What Kelly seems to miss, though, is the fact that these exoselves will be operated by commercial entities. And these entities have a history of embarking on infamous “enshittification” journeys (case in point: Meta now charges for essential features of its AR glasses). And as such, the relationship we have with our AI might not be something the user chooses, but platforms – not users – will engineer which of the four types you get, tuned to engagement or subscription revenue rather than your interests.

In which case, Kelly’s little admission at the very end of his piece might be the least of our worries (emphasis mine):

This second self will demand a new kind of relationship, one we haven’t had before — and its immense benefits will arrive bundled with immense problems. Every ailment that afflicts our born self will likely show up in the exoself too, plus novel ones we haven’t seen yet. Learning to use an exoself wisely will be one of the major lessons of a life lived this way. It will take years before society works out anything like best practices — we’re still working on those for social media. There will be multiple models and personality types to choose from. And there will be heart-wrenching stories of people losing their exoself — the worst case being simply that the platform went out of business.

↗ Link

By now, fearmongering about AI has become the norm. From OpenAI and Anthropic warning the public about the insane danger their latest frontier models pose, to AI hypers and doomers taking positions on either end of the extreme to take the headlines. It’s all rather annoying and tiresome. And just when you think you have seen it all, along comes this (emphasis mine):

“AI is a global technology with global benefits, global harms, and a consistent tendency for new capabilities to eventually proliferate," Stephen Casper, a computer scientist at MIT who spoke at a major AI conference in Beijing this month, told Wired. “One thing that almost everyone in AI can agree on right now is that AI doesn’t need a Chernobyl moment,” he added. Casper didn’t elaborate further on this analogy, but by invoking the infamous nuclear disaster, it’s clear that the fear isn’t just over the catastrophe itself.

Yeah, we don’t need a Chernobyl moment. Also – whatever that means. Jeez!

↗ Link

Say what you want about the leaders of AI shops Anthropic and OpenAI, criticize their truly dumb game of “our models are so good and powerful, they need to be locked up” fearmongering marketing play, but what happens now, with a national government deciding nilly-willy who gets access to AI, is weird, wrong, and plain dangerous.

The Trump administration is requiring both Anthropic and OpenAI to get approval for each new customer of their most powerful AI technology.

Assume (and that is still, to a degree, an assumption), for a moment, that the difference between frontier models and the rest of the bunch does matter in terms of economic impact (firms with access to frontier models have an actual competitive advantage), and you realize how messed up it is that the US government is shifting the playing field in very real ways – not just on a firm-by-firm basis, but for nation-states and whole regions. But then, of course, all this might not matter all that much as open-source models such as GLM-5.2 have become (allegedly) as good as (or better than) Mythos-class models. Maybe the genie is out of the bottle…

↗ Link

Last summer, Stanford professor Erik Brynjolfsson published a comprehensive (albeit early) study on the impact of AI on the job market. Crunching a large data set made available through payroll provider ADP, he found that the impact of AI on jobs is real and measurable. Now he is back with more, and more current, data. And it doesn’t look any prettier.

For workers ages 22 to 25, employment in highly AI-exposed occupations is now shrinking at 3.8% per year and the early-career decline sharpened after year one — 2.8% decrease to April 2024, growing to a more than 4% decline per year since. The average decline on a month-to-month basis averages about −0.3% but Brynjolfsson notes that trend is noisy, compared to the year-over-year deceleration.

Link above is to the full study – it’s worth digging into the data to get the full picture.

↗ Link

Much (much!) has been written about the pros and cons of remote work. Companies left and right have been requiring people to come back to the office – or embrace the remote/hybrid model. Studies have been done to show that people are happier when they can work from home; others show that remote work has a measurable impact on creativity and innovation across the firm. But, until now, little has been written (let alone studied) about the impact of remote work on mental health. A recently published study in Science looked at a large swath of data from studies done on remote work (note that this is US-only data) and came to a sobering conclusion (emphasis mine):

After the pandemic, workers in remote-capable jobs spent more time working alone and avoided social activities with their friends, remaining more isolated both during and after work. This pattern was most pronounced among remote workers living alone: They spent entire days without human contact and their mental distress, use of mental healthcare, and antidepressants increased acutely.

It’s not a pretty picture and warrants further study, as well as consideration on behalf of employers (outside of the ROI debate).

↗ Link

When Jeffrey and I were at Singularity University, we used to comment on the fact that the average plumber in Silicon Valley made about as much (and sometimes more) money than the average software engineer. The same was (and is) true for many other trades. Plus – try to get a plumber to come to your house if you live in the Bay Area and you’ll experience first hand how hard it is to even find one. Nvidia’s CEO Jensen Huang just made the same point:

“If you’re an electrician, you’re a plumber, a carpenter—we’re going to need hundreds of thousands of them to build all of these factories,” Huang told Channel 4 News in the U.K. in late 2025. “The skilled craft segment of every economy is going to see a boom. You’ve going to have to be doubling and doubling and doubling every single year.”

Aside from him potentially being right – it’s kinda sad to think we live in a world where the value of human labor is to keep the machines running. Brave New World indeed.

↗ Link

I do disagree with the multi-level marketing analogy (as MLMs live and die by their pyramid scheme nature of recruiting new people into the system), but Matthew Hughes does have a point when it comes to the notion that the “vibe-code yourself to millions” message of too many of the GenAI startups is predatory, sad, and dangerous. It makes me wonder how we will remember this moment in time.

And I am concerned that the fear I’ve described is being exploited by companies like Replit and Cursor (which is also doing the exact same influencer marketing schtick, albeit not as aggressively as Replit), who are touting their services as a way for people to escape the precariousness of this current moment.

Hughes’s best point is that vibe-coding is worse than Herbalife because at least Herbalife tells you the price, whereas vibe-coding tools notoriously burn through tokens with little price transparency. Which is also why it isn’t Herbalife.

↗ Link

The newest student AI cheating tools don’t just write the essay for you – they slow-type it into Google Docs, typos and corrections and all, to slip past the detectors, while the human does nothing of the sort. We’ve now built software whose entire job is to fake the effort of learning, because effort is the thing we grade.

To me, this is the whole story; not the cheating per se. Cheating in school is as old as school itself (heck, I am guilty as charged here). The truly interesting question is why a student who’s just signed up for years of student loan debt would so cheerfully skip the part they’re actually paying for. The answer might be that they’re not being lazy (or dumb), they’re being rational. Bryan Caplan made the case years ago in The Case Against Education: most of what a degree buys you isn’t the knowledge, it’s the signal. If the credential is the product and the learning is optional, then handing the work to a bot that types its own typos is exactly what a smart customer does. Yes, this is the calculator panic all over again – except the calculator never pretended to be you doing the arithmetic. This one does. Which leaves one uncomfortable question for every school grading the take-home essay: what did we think we were measuring all along?

↗ Link

The prevailing narrative at the moment is that AI is coming hard for software engineering jobs (from mass layoffs to recent graduates not being able to find a job and everything in between). Turns out, the data doesn’t actually support this narrative. Arvind Narayanan and Sayash Kapoor explored this topic in a thoughtful post on their “AI as normal technology” Substack. The reason is what Narayanan and Kapoor call the “decide-execute-deliver sandwich” – of which AI compresses the “execute” part but doesn’t budge on the other two.

Across 100,000 developers on GitHub, the researchers found that AI agents led to an eight-fold increase in the number of lines of code written, consistent with the idea that AI almost completely compresses the Execute layer of the sandwich. But this led to only 30% more releases, strongly suggesting that human bottlenecks (the Decide and Deliver layers) remain in place.

In my eyes, this has pretty far-reaching implications for other professions – as the two authors also point out:

In this essay, we argue that there is enough evidence to reject the narrative that once AI capabilities reach a certain threshold, it will cause mass layoffs. Given that this is true even in a sector with very few regulatory barriers, most other professions are likely to be even more cushioned.

Highly recommended reading.

↗ Link

This is pretty funny – Tim Ferriss (of “The 4-Hour Workweek” fame) writes a long post about “Has AI Already Killed How-To Nonfiction? Sales Trends, My Personal Data, and What It Might Mean for the Future.” It has some interesting insights and data points, and you might want to read it. But… it is also, likely, written (or co-written) by AI… ;)

Some telltales (unless Tim has adapted his writing style to sound like AI now):

My head has been spinning after getting a spreadsheet roughly a week ago.

But, let’s be honest: one quarter doesn’t make a trend. So let’s zoom out and look at my full catalog over a few years.

And my personal favorite:

Let that sink in for a minute.

The problem (for me at least) isn’t that Tim is (or isn’t – who knows, maybe his writing just sounds like this) using AI to write his blog posts – it’s that, due to the fact that I must have read the “Let that sink in for a minute.” line a million times by now (as it’s a staple of AI-generated slop), I am just so much less engaged with his article. Which is a shame – as it does make a good point (or so I believe, as I couldn’t get myself to do more than skim it…).

↗ Link

On her blog, German pharmaceutical database specialist Ava outlines the creative uses her company found for AI – it makes you cringe, smirk, and realize that this exact thing might be going on in many companies the world over.

We have recurring house-wide meetings where groups are asked to show off their LLM projects. They register them, try them out for a couple months, and then come back presenting their results. I have attended all of these meetings so far, and there was not a single one that actually worked out. All projects ended with the conclusion that this isn’t workable, that this isn’t saving time, or that it over-complicates things. Hundreds of people, different teams, people enthusiastic about AI, all kinds of projects, and there wasn’t a single success.

For one, it was shown that you can ask the bot how it feels today. That wasn’t presented as a joke, or being sarcastic; no, it was shown very seriously, I guess under the guise of how cool and futuristic and human it is. […] Next up was the great use case of downloading the cafeteria menu (which is a 1 page nicely designed Excel sheet, like a timetable, showing the different options for each day) from the intranet, giving it to ChatGPT, and asking it what’s for lunch on Wednesday.

Read the whole thing. It’s gold. And then, maybe, compare it to your reality in the company you work for… ;)

↗ Link

Classroom teachers in the US are becoming increasingly worried about the impact of AI on their students – specifically their critical thinking skills (which, ironically, are the very skills that are most needed in an age of AI).

Christa Corricelli, a special education teacher at Saugus Middle/High School outside Boston, says AI could be a valuable technology for learning, but too often students are using it as an answer machine – not a tool to bolster their thinking. “I think students who aren’t already intrinsically self-motivated to be critical thinkers, like that top 1% of the class … I think people who are not already that personality type, we’re going to see those critical thinking skills atrophy over time,” Corricelli says.

We likely won’t know how the use of AI, both on the teacher and student side, will impact our children’s mental and social development, nor what this will mean for the world at large – but it is surely something we might want to try to get right before it’s too late.

↗ Link

The use of generative AI in genetic research, so far, has been held back by the complicated nature of turning DNA sequences into their physical forms – simply stated: The machines can dream up new DNA sequences much faster than we are able to assemble (and then test) them. This is about to change:

The technique, called Sidewinder, can assemble dozens of genetic sequences simultaneously in a single test tube, producing just one incorrect junction for every 10 million assembly events – a level of precision that far surpasses conventional methods, which misfire roughly once every 10 to 30 joins. Sidewinder also draws on cheap raw materials that have until now been too difficult to use reliably.

and:

In a demonstration of how squarely Sidewinder targets this bottleneck, the team behind the technique, led by Caltech synthetic biologist Kaihang Wang, harnessed the power of Evo 2 to redesign a 12,500-letter DNA sequence of the E. coli genome in silico and then used Sidewinder to build it from scratch – with no errors. Sequences of that length can encode entire biochemical pathways, laying the groundwork for engineered microbes that manufacture drugs, biofuels, or specialty chemicals, and eventually to the assembly of vast DNA constructs approaching complete artificial genomes.

Now, just as a thought experiment, consider what this could mean for the hotly-debated issue with the capability of frontier models to be use to assist with biological warfare. With great powers comes great responsibility…

↗ Link

A new study from MIT’s Media Lab shows the correlation of tool use (in three groups: ChatGPT, Google Search, and no tool use) with brain activity. With the caveat that this is a small study, the results are not pretty:

EEG analysis presented robust evidence that LLM, Search Engine and Brain-only groups had significantly different neural connectivity patterns, reflecting divergent cognitive strategies. Brain connectivity systematically scaled down with the amount of external support: the Brain‑only group exhibited the strongest, widest‑ranging networks, Search Engine group showed intermediate engagement, and LLM assistance elicited the weakest overall coupling.

As the study concludes: “We demonstrate the pressing matter of exploring a possible decrease in learning skills.” Ouch.

↗ Link

Add this to your AI datacenter bingo card: not only are GPUs and RAM in short supply, energy is a major limiting factor, and water consumption is a massive concern; we also don’t have enough electricians to build out the infrastructure.

That said, the Lake Mariner site is about one hour away from the Buffalo Bills’ stadium, which delivers another benefit. The completion of a massive refurb at the Bills has freed up hundreds of electrical contractors. And it is trades, specifically electricians, which are the biggest bottle neck for datacenter projects, said Farrell.

↗ Link

Scale Labs just updated their Remote Labor Index (RLI) – a measure of how well AI agents are actually able to do work in the real world (“Evaluating the capability of AI agents to perform real-world, economically valuable remote work”). The tl;dr:

Absolute Automation is Near Zero: Current agents perform near the floor. At the time this leaderboard was launched, the highest-performing agent (Manus) achieved a 2.5% automation rate, with other models performing worse. This indicates systems fail to complete the vast majority of projects to a professional, client-ready standard.

The upshot? They are getting better.

↗ Link