Friday Wrap: Volume One
A round-up of AI-related news from the web.
Welcome to this new, regular (?) feature : The Friday Wrap. This is a kind of sub-newsletter that I'll be doing each week covering random things that caught my attention. Some of those will be links to other people's work on the web and some will be the transient thoughts that pass through my head without ever solidifying into a full-blown post. This week, they're almost exclusively about artificial intelligence which isn't particularly by design, just a reflection of the world's current obsession with AI.
LLMs are not models of the world
Large language models are just that: models of language. Gary Marcus discusses at length how this is different to a model of reality and how this discrepancy explains, amongst other things, ChatGPT losing at a game of chess played against an Atari 2600 from 1977. Marcus on AI: Generative AI’s crippling and widespread failure to induce robust models of the world
Meanwhile, researchers at a collection of US universities have coined the phrase "potemkin understanding" to refer to the semblance of understanding that LLMs exhibit on the surface but fail to demonstrate deep down. AI models just don't understand what they're talking about
It's models all the way down
What's better than using an LLM? Using several LLMs! And how do you choose which LLM to use for which query? Using another machine learning model! LLM routers exist to direct particular queries to particular models based on user-defined characteristics such as which is faster, cheaper or better tuned to that particular task. Now researchers have developed Arch-Router, a 1.5 billion parameter neural net model that, it's claimed, learns user preferences for different models. Arch-Router
Faster, better?
One of many time-saving hacks that's claimed to optimise your learning is playing lecture recordings, podcasts and videos at double (or greater) speed. A cognitive scientist has looked into the effects of this and assesses whether it really helps or hinders a learner's progress. What happens to your brain when you watch videos online at faster speeds than normal
AI, take my job, please!
A common concern recently is that AI is going to take away people's jobs. At the Financial Times, statistician Tim Harford dives in to the debate. He distinguishes between jobs that require skilled task completion and those that also require judgement and problem-solving abilities. He suggests that the latter group are those that are safe from AI in the same way that the profession of accountancy survived the introduction of the spreadsheet because it was always about more than just the calculations. Whose job is safe from AI?
Meanwhile, the sky is falling for computer science graduates, according to The Times. It claims that the brightest graduates from top programmes at MIT and Stanford are unable to get entry-level jobs because they're all being done by AI. An introduction to correlation and causation might go down well at The Times. Sure, the job market for new graduates is tough at the moment but there are a multitude of reasons for that, most of them related to the global economy rather than the verbiage-spilling machines of AI. ‘I have two degrees in computer science. AI is stealing my jobs’
AI is meta and meta is AI
Mark Zuckerberg is off on one again, according to Ars Technica. A few years after raving about the "metaverse" (so good they renamed the company after it), Zuck is now evangelising the wonders of "AI superintelligence". It's going to be the future of the internet, apparently. Just like the metaverse wasn't. Meta’s “AI superintelligence” effort sounds just like its failed “metaverse”
And finally
Hacker News has been good to this newsletter, sending decent amounts of traffic to several of my articles. So it's intended in good humour that I link to HN Slop - Josh Cohenzadeh's tongue-in-cheek AI-generated startup ideas derived from the current front page of Hacker News. Current top-rated idea: SwearySkyscraper, "a startup that develops novel and personalised swear words to help people better cope with pain, and integrates these swear words into a liquid damping system to stabilise skyscrapers during earthquakes".
