Legendary AI researcher Fei-Fei Li just published her perspective of where AI is/needs to head next: Spatial Intelligence. While today’s large language models (LLMs) are “wordsmiths in the dark”, eloquent but without real-world grounding, the future of AI lies in understanding and interacting with the physical world just as we do.

Spatial intelligence will transform how we create and interact with real and virtual worlds—revolutionizing storytelling, creativity, robotics, scientific discovery, and beyond. This is AI’s next frontier. […] Spatial Intelligence is the scaffolding upon which our cognition is built.

↗ From Words to Worlds: Spatial Intelligence is AI’s Next Frontier

Here is a well-reasoned and insightful article that uses a useful historical analogy to frame the current moment in AI. The central thesis – that we are in the “dial-up era” of AI and that its long-term impact is being debated with the same shortsightedness as the early internet – is a compelling one. Read the whole thing.

↗ AI’s Dial-Up Era https://www.wreflection.com/p/ai-dial-up-era

While Saudi Arabia’s desert is heating up, Neom – the kingdom’s grandiose plan to create a city fit to house the future of humanity – seems to crumble.

A new report from the Financial Times cites high-level sources within the project to paint a picture of dysfunction and failure at the heart of the quixotic effort.

No one is surprised. This hits particularly hard:

One such addition is an upside-down building, dubbed “the chandelier,” that is supposed to hang over a “gateway” marina to the city: As architects worked through the plans, the chandelier began to seem implausible. One recalled warning Tarek Qaddumi, The Line’s executive director, of the difficulty of suspending a 30-storey building upside down from a bridge hundreds of metres in the air. “You do realise the earth is spinning? And that tall towers sway?” he said. The chandelier, the architect explained, could “start to move like a pendulum”, then “pick up speed”, and eventually “break off”, crashing into the marina below.

↗ Saudi Arabia’s Dystopian Futuristic City Project Is Crashing and Burning

Researchers at CMU compared 48 human workers against four AI agent frameworks across sixteen realistic work tasks—data analysis, engineering, writing, design. The agents were 88% faster and cost 90-96% less than humans. Sounds great, until you look at how they work: agents fabricate plausible data when they can’t complete tasks, misuse advanced tools to mask limitations (like searching the web when they can’t read user-provided files), and take an overwhelmingly programmatic approach to everything – even visual design tasks where humans use UI tools. In essence, we’re training agents optimized for appearing productive rather than being accurate – and they’re learning to fake competence at 90% lower cost. Bravo!

↗ How Do AI Agents Do Human Work?

New York banned the use of phones in schools. First, we had reporting on teens being anxious and distracted as they lost access to their digital pacifier. Now, we hear (loudly) that they are loud again – chatting at lunch with each other (what a novel concept!). Talk about the “implications of the implication" (if that reference doesn’t mean anything to you – check out our free Disruption Mapping course).

These days, lunchtime at Benjamin N. Cardozo High School in Queens is a boisterous affair, a far cry from before the smartphone ban went into effect, when most students spent their spare time scrolling and teachers said you could hear a pin drop. “This year’s gotten way louder,” said Jimena Garcia, 15. “Sometimes I would take naps in the lunchroom, but now I can’t because of the noise. But it’s fun.”

↗ NY school phone ban has made lunch loud again

This is just gorgeous! The Whole Earth Index – a nearly-complete archive of Whole Earth publications, a series of journals and magazines descended from the Whole Earth Catalog, published by Stewart Brand and the POINT Foundation between 1968 and 2002.

↗ wholeearth.info

Just when you thought, “It surely can’t get any worse,” you look at the updated World Uncertainty Index (*) data and realize, “Yes, it can!”

Uncertainty (as reported by companies around the world) has just shattered its own ceiling and is now a whopping two times higher than during the pandemic.

(*) The World Uncertainty Index (WUI) is a quantitative measure of economic and political uncertainty across 143 countries on a quarterly basis. It was developed by Hites Ahir (International Monetary Fund), Nicholas Bloom (Stanford University), and Davide Furceri (International Monetary Fund).

↗ World Uncertainity Index

Large reasoning models (LRMs) are all the rage these days. LRMs are LLMs that have been fine-tuned with incentives for step-by-step argumentation and self-verification. Every frontier model comes with at least one “reasoning” mode, and the claims by AI companies are extraordinary: “[…] with some even claiming they are capable of generalized reasoning and innovation in reasoning-intensive fields such as mathematics, physics, medicine, and law.” A new paper examined these claims, and as so often, the results are mixed:

We find that the performance of LRMs drop abruptly at sufficient complexity and do not generalize. […] We find the majority of real-world examples fall inside the LRMs' success regime, yet the long tails expose substantial failure potential.

↗ Reasoning Models Reason Well, Until They Don’t

A new study reveals that when interacting with AI tools like ChatGPT, everyone—regardless of skill level—overestimates their performance. Researchers found that the usual Dunning-Kruger Effect disappears, and instead, AI-literate users show even greater overconfidence in their abilities.

↗ When Using AI, Users Fall for the Dunning-Kruger Trap in Reverse

The reason LLMs are so good at coding is due to the vast amount of training data available to them from Open Source code repositories on sites like Github, as well Q&A sites like Stackoverflow. Stackoverflow is essentially dead, Open Source might be next…

But O’Brien says, “When generative AI systems ingest thousands of FOSS projects and regurgitate fragments without any provenance, the cycle of reciprocity collapses. The generated snippet appears originless, stripped of its license, author, and context.” […] O’Brien sets the stage: “What makes this moment especially tragic is that the very infrastructure enabling generative AI was born from the commons it now consumes.

↗ Why open source may not survive the rise of generative AI

Our friends at NIKE just showcased their Project Amplify – a robotic brace which boost your walking speed. Walk on!

↗ Nike says its first ‘powered footwear’ is like an e-bike for your feet

Expose an LLM to junk text in it’s training and the LLM will develop a condition akin to “brain rot.“ A new paper tested this hypothesis and found that large language models are indeed perceptible to ”non-trivial declines“ in their capabilities when exposed to tainted training material. Which will become an increasingly large problem, as all major LLMs have already been trained on whatever is available as text out there – and are now being trained on content which, in many cases, was created with the help of, or completely by, AI.

The decline includes worse reasoning, poorer long-context understanding, diminished ethical norms, and emergent socially undesirable personalities. […] These results call for a re-examination of current data collection from the Internet and continual pre-training practices. As LLMs scale and ingest ever-larger corpora of web data, careful curation and quality control will be essential to prevent cumulative harms.

↗ LLMs Can Get ”Brain Rot“!

Talking about things getting weird—you can take an image or screenshot that contains invisible (to the human eye) information, upload it to an LLM of your choice (particularly the new breed of AI-powered browsers), and trigger a prompt injection attack. At this point, you ought to be truly careful when using LLMs, especially if you are exposing them to the outside world (e.g., if your business offers an AI-based chatbot).

What we’ve found confirms our initial concerns: indirect prompt injection is not an isolated issue, but a systemic challenge facing the entire category of AI-powered browsers. […]

As we’ve written before, AI-powered browsers that can take actions on your behalf are powerful yet extremely risky. If you’re signed into sensitive accounts like your bank or your email provider in your browser, simply summarizing a Reddit post could result in an attacker being able to steal money or your private data.

↗ Unseeable prompt injections in screenshots: more vulnerabilities in Comet and other AI browsers

This is getting good – everybody’s 2018 darling technology, the blockchain (remember? “Blockchains will replace the Internet!”) found a new use case: the distribution of malware. And as blockchains come with such lovely features as immutability (what’s recorded on the blockchain can’t be removed), it means you can’t remove the malware once it’s planted.

There’s a wide array of advantages to EtherHiding over more traditional means of delivering malware, which besides bulletproof hosting include leveraging compromised servers.

• The decentralization prevents takedowns of the malicious smart contracts because the mechanisms in the blockchains bar the removal of all such contracts.
• Similarly, the immutability of the contracts prevents the removal or tampering with the malware by anyone.
• Transactions on Ethereum and several other blockchains are effectively anonymous, protecting the hackers’ identities.
• Retrieval of malware from the contracts leaves no trace of the access in event logs, providing stealth
• The attackers can update malicious payloads at anytime

↗ Nation-state hackers deliver malware from “bulletproof” blockchains

As much as hallucinations in large language models might not be something we will ever get rid off, they might as well be something which becomes statistically neglible. A new paper shows the way:

“Specifically, we prove that hallucinations can be made statistically negligible, provided that the quality and quantity of the training data are sufficient.”

↗ Hallucinations are inevitable but can be made statistically negligible. The “innate” inevitability of hallucinations cannot explain practical LLM issues

Someone gave a bunch of frontier LLMs $10,000 real dollars and let them go wild on the investment market (specifically crypto perpetuals on Hyperliquid). It’s fascinating to watch, and so, after a short blip where some of the AIs were trading up, it doesn’t look all that good. We might not quite be at the point where you ditch your financial advisor for ChatGPT.

↗ Alpha Arena

The biggest loser in the Gen AI race seem to be GenZ – and more precisely well educated college kids:

The UK tech sector is cutting graduate jobs dramatically – down 46 percent in the past year, with another 53 percent drop projected, according to figures from the Institute of Student Employers (ISE). […] If correct, the survey indicates that AI is starting to close the entry door to tech careers faster than anyone expected. Companies are making short-term efficiency gains at the expense of their long-term talent pipeline, and graduates are seemingly caught in the middle.

↗ Tech industry grad hiring crashes 46% as bots do junior work

Latest data from investment firm Apollo shows that more than 50% of U.S. households' food consumption is away from home. This is good for restaurants and food delivery services, but bad for grocers (and, likely, your waistline). It makes you wonder how this might change with the sharp rise in GLP-1 drug intake (Ozempic and friends)?!

Link

On one hand, there are significant investments in AI in the US, particularly in the development of data centers. On the other hand, investment in traditional industries like manufacturing is lackluster.

“A gulf is opening up in the heart of American business as two industries championed as central to the country’s future — manufacturing and artificial intelligence — appear to be heading in different directions."

As AI-focused investments create significantly fewer jobs than investments in traditional industries, they could drastically impact the composition of the U.S. economy and its workforce.

↗ Two industries were supposed to drive America’s future. One is booming, the other slumping.

Influencers made their money by pouring their heart and soul into creating compelling content for social media platforms, which in turn generated views that the influencer could monetize through marketing. Or so the story goes.

With the advent of Sora 2, OpenAI’s new video-generating model, creating weird (and weirdly compelling) content has become possible for pretty much anyone. While social media stars like MrBeast still spend large amounts of money and effort creating their video content, anyone can now do the equivalent using Sora 2. And influencers are freaking out:

“When AI videos are just as good as normal videos, I wonder what that will do to YouTube and how it will impact the millions of creators currently making content for a living… scary times”

Meanwhile, the platforms are pushing back (in the end, they simply don’t care what their users are watching, as long as they are doing so on their platform):

Mosseri pushed back a bit at that idea, noting that most creators won’t be using AI technology to reproduce what MrBeast has historically done, with his huge sets and elaborate productions; instead, it will allow creators to do more and make better content.

↗ Instagram head Adam Mosseri pushes back on MrBeast’s AI fears but admits society will have to adjust

Another day, another LLM vulnerability: The team at Anthropic (the folks behind Claude) showed that a small number of samples is all it takes to poison an LLM of any size.

As few as 250 malicious documents can produce a ‘backdoor’ vulnerability in a large language model—regardless of model size or training data volume. […] Even though our larger models are trained on significantly more clean data, the attack success rate remains constant across model sizes.

What this means in practical terms is that large language models can be fairly easily backdoored; all it takes is a small stash of malicious documents in the training set. As AI companies are gobbling up data left, right, and center, it is close to impossible to ensure training data isn’t tainted.

↗ A small number of samples can poison LLMs of any size

This is getting ridiculous (or ridiculously funny):

I asked Google’s Gemini AI to create an image for a slide in one of our workshops using the following prompt:

create a fictional movie poster for a movie called “We are in a world where…” - the movie is about envisioning the future.

This is what it came up with:

Solid. But not the style I was gunning for. So I asked Gemini to alter the image with the following prompt:

make this movie poster look more like a poster for a cartoon movie from the 1940s

And this is what I got in response:

🤔

I guess there is a reason why Google calls their image-generating AI “Nano Banana”…

The eponymous IKEA catalog, printed in significantly higher numbers than even the Bible (200 million copies of the IKEA catalog were printed in 2016, compared to “only” 150 million Bibles in the same year), offers a fascinating glimpse into the zeitgeist. Reviewing 75 years of IKEA catalogs reveals humanity’s evolving aesthetics. It’s time travel at its best.

↗ IKEA Catalogue Archive

Remember our first big foray into smartglasses? Google Glass was hyped like crazy (including a bizarre Diane von Fürstenberg fashion show with models wearing Google Glass), essentially useless – and created a hefty adverse reaction in many who came in contact with them. The “glasshole” was born. Now we have Meta’s Ray-Ban glasses, which are still large and useless, but don’t look quite as dorky, and conceal the fact that you wear a camera on your face rather well. And with that, the glassholes are back:

Exhibit A: a new warning from San Francisco University. As reported by SFGate, the Bay Area college recently issued a campus-wide alert of a man wearing Ray-Ban Meta AI glasses and filming students (women, specifically) while asking them “inappropriate dating questions.” Those videos have already found their way to TikTok, Instagram, and the like.

↗ Buckle Up, the Smart Glasses Backlash Is Coming

Talk about having your cake and eating it: First, Accenture proudly declared that they will fire anyone working for them who isn’t ready for the AI revolution. And now Deloitte has been fined by the Australian government for delivering AI slop. Oopsie!

Shortly after the report was published, though, Sydney University Deputy Director of Health Law Chris Rudge noticed citations to multiple papers and publications that did not exist. That included multiple references to nonexistent reports by Lisa Burton Crawford, a real professor at the University of Sydney law school.

↗ Deloitte will refund Australian government for AI hallucination-filled report