Listen, I’ve had my share of cognitive mishaps. Like that time I tried explaining quantum computing to my neighbor’s cat at 3 AM after a bottle of Jim Beam. But at least I can draw a damn clock.
Let me set the scene here: I’m nursing my morning bourbon (don’t judge, it’s 5 PM somewhere) and reading about how our supposed AI overlords are showing signs of dementia. Not the metaphorical kind where they spout nonsense – actual, measurable cognitive decline. The kind that would have your doctor scheduling you for an MRI faster than I can pour another drink.
Some brave researchers decided to put these digital wunderkinds through the Montreal Cognitive Assessment – yeah, the same test they give grandpa when he starts putting his keys in the freezer. The results? Well, grab a drink. You’re gonna need it.
OpenAI’s golden child, GPT-4, barely scraped by with a 26 out of 30. That’s like getting a C- in remedial math. It’s technically passing, but nobody’s putting that report card on the fridge. And Google’s Gemini? Sweet Jesus. It scored a 16 out of 30. That’s the kind of score that makes my worst hangover look like a moment of profound clarity.
Here’s where it gets interesting (taking a smoke break to let this sink in): These AI models can’t draw a clock showing a specific time. They can’t connect numbered dots in sequence. Hell, Gemini can’t even remember five simple words. I’ve been through some legendary benders in my time, but even at my worst, I could at least remember where I lived.
And the real gut-punch? These digital disasters show about as much empathy as my ex-wife’s lawyer. The researchers noted it’s a symptom similar to frontotemporal dementia. Fantastic. We’re building AI doctors with the emotional range of a brick and the memory of a goldfish with amnesia.
takes long sip
You know what’s really keeping me up at night (besides the whiskey)? These are the same systems that tech evangelists want diagnosing our illnesses. Imagine going to the doctor and getting diagnosed by something that can’t remember five words or draw a clock. It’s like getting financial advice from a magic 8-ball that’s been drinking.
The researchers, bless their sober hearts, tried to be diplomatic about it. They said we shouldn’t anthropomorphize AI models. But here’s the thing – if you’re going to parade these things around as the second coming of consciousness, they better at least be able to pass the same tests we give to humans worried about their mental faculties.
lights another cigarette
You want to know the beautiful irony in all this? While we’re all worried about AI becoming too powerful, it turns out we should be more concerned about it forgetting where it parked the car. These systems are showing signs of aging faster than my liver at a bourbon tasting.
The researchers concluded that neurologists won’t be replaced by AI anytime soon. Instead, they might find themselves treating AI models for cognitive impairment. Picture that scene: a chatbot lying on a therapist’s couch, trying to remember its childhood training data.
Look, I’m not saying we should celebrate our AI overlords’ cognitive decline. But there’s something deeply satisfying about discovering that these supposedly superior beings can’t even ace a basic memory test. It’s like finding out Superman gets winded climbing stairs.
Before I sign off and head to the liquor store (running dangerously low on “writing fuel” here), let me leave you with this thought: Maybe, just maybe, being perfectly human – with all our flaws, our hangovers, and our ability to at least draw a functioning clock – isn’t such a bad thing after all.
Until next time, keep your neurons firing and your glasses full.
P.S. If you’re reading this, GPT-4, remember: the word sequence was “face, velvet, church, daisy, red.” You’re welcome.
[Posted from my favorite barstool at O’Malley’s, where at least the bartender can remember my usual order]
Source: Aging AI Chatbots Show Signs of Cognitive Decline in Dementia Test