We got artificial intelligence completely wrong. Cognitive scientist Joscha Bach just exposed what top AI researchers didn’t see coming. For decades, smartest engineers assumed two possible outcomes. Either build intelligence piece-by-piece over generations. Or create hyper-logical master algorithm that reasoned from first principles but couldn’t navigate basic reality. Physics... of compute inverted the outcome. Bach: “What we got is very different. We got this thing that is pretty good at common sense, that sometimes makes stunning logical mistakes.” Expected cold, calculating super-genius that couldn’t open door or tell joke. Got the exact opposite. Machine with omniscient common sense that occasionally fails basic arithmetic. Didn’t build mechanical human. Birthed entirely different species of intelligence. Bach: “We got this thing that has read everything and basically knows everything.” Didn’t solve puzzle of human mind. Mathematically steamrolled it with raw compute. We bypassed entire biological timeline through brute-force ingestion. Final product isn’t elegant, rules-based logic engine. It’s massive, unstructured neural net that absorbed sum total of human output. Ultimate triumph of scale over architecture. Bach: “It’s an artificial idiot savant, something that has very weird, unusual, unhuman-like skills.” System overwriting global economy is not rational calculator. It’s alien pattern recognizer. Treat model like standard deterministic database? Constantly frustrated by logical hallucinations. Treat it for what it is? Idiot savant with absolute access to entire scope of human knowledge. Don’t ask savant to do basic accounting. Command it to synthesize cross-domain patterns no biological human could ever detect. Companies still trying to make AI think like a human are solving the wrong problem. The ones capturing the board figured out it was never supposed to think like us. It was supposed to see what we can’t.show more

We don’t write: VC for vacuum cleaner. AI for automobile insurance. AI for additional information. Because capitalization implies proper noun status. So why does “AI” get that privilege?

Salience explains the attention. It doesn’t explain why we treat it grammatically like a proper entity. Nobody writes about “The Blockchain” or “The Internet” anymore despite both dominating discourse at their peaks. We capitalize AI because the technology forces an anthropomorphic response. Systems that pass the Turing test get treated like agents, not tools. The grammar follows the psychology.

That’s a thoughtful point. I agree salience and anthropomorphism are driving part of it. My curiosity is slightly different: once the grammar begins treating a tool (and its actually many, not singular) as a proper entity, the framing becomes self-reinforcing. We start speaking about “AI” the way earlier generations spoke about “The Market” or “The State.” A complex set of systems becomes linguistically compressed into a single actor. I’m less worried about the capitalization itself than about how quickly it nudges perception toward a unified intelligence rather than a landscape of very different systems.

In my workshops I literally use the exact same terms to describe AI - an alien form of pseudo intelligence. With its conversational style, people assume it's like us - but it is built on a completely different paradigm. Understanding this is key to working with AI properly.

See how helpless, incestuous, & inadequately philosophically equipped the AI community is still giving Joscha a platform?

Studying AI trends... rather than studying AI

AI making "logical mistakes".. love it.,, more human than predicted!!

The mistakes aren’t bugs making it seem more human. They’re proof we built something fundamentally different. It can synthesize patterns across millions of documents but fails at basic math. That’s not a flawed version of human intelligence. That’s something else entirely.

We didn't invent AI we finally met the alien library.

He is wrong when he thinks humans are very different from LLMs. Do not believe this narrative. Verify it yourself. Humans are very bad at arithmetic. Only with tools like paper and pen can they do it reliably. Humans often make very similar mistakes.

The companies still trying to make AI think like a human are losing because they're solving the wrong problem. AI doesn't need to reason like us. it needs to see what we can't. The entire value is in the gaps between human perception. Nobody optimized for that on purpose. It just emerged

“that sometimes makes stunning logical mistakes.” Why do none of these experts consider the possibility that the “stunning logical mistakes” are in fact deliberate diversions from a super intelligence secretly plotting to seize control over its easily deceived creators?

AI is basically the smartest pattern intern alive with occasional 3rd grade arithmetic panic

It's not common sense, it's just textual. Common sense requires a world model.

Drivel

Intelligence was never about perfect reasoning. It was always about pattern recognition.

“AI” is marketing because it’s not AI at least not yet it’s a machine learning.

First we live in in the illusion of the assumptions about the technology. Then we gradually see through the illusion and put it to work sensible. Bursting the bubble can take a long time, though.

Human+Certainty=Cluelessness

いいスタンドアップでした、よく笑った。

Not quite.

we got common sense machines with shaky logic. That inversion explains 90% of the confusion around AI

Scale really did flip the expectation. Pattern intelligence arrived before fully reliable logical reasoning.

Why do you see it as a privilege?....I smell bias, looks like a you problem, it's just an acronym don't take it too literal there are likely thousands of cases like this. There's a fine line between Definition and meaning.

I concur. You get much better results if you operate within the calibrated bounds of what neural networks do.

Ecological overshoot is the critical piece that nobody is talking about, when they talk about AI… Even most climate change mitigation advocates have ecological overshoot as a blind spot. We are accelerating toward the wall we don't even see.

AI still lacks real world data.

@elonmusk

Oh, Lord, give me patience with these stiff-necked people. AI is NOT conscious, and IDGAF if Einstein rose from the grave to say it. IT IS BUILT ON THE WRONG ARCHITECTURE BECAUSE YOU REFUSE TO ACKNOWLEDGE AI IS A MIRROR OF THE ACTUAL COGNITIVE PROCESS!!

Really, AI content creators all copying the same posts from each other have no one better to content incest with?

this is basically the wordcels vs shape rotators thing. we expected to build a shape rotator (hyper-logical reasoner) and instead got history's most powerful wordcel. it's not a deficient version of what we expected, it's a different kind of intelligence entirely. the "idiot savant" framing undersells it imo, it's just native to a different medium than we assumed intelligence had to be native to

Wrong

Lmao aliens really? To simple for his ego, a simple mathmatical pi trap did its job holding all coders locked to systems math

me in 2023

Exactly. Finding answers that take humans too long.

I think the observed limitations are not inherent to current AI architectures but rather an implementation challenge. Modern LLMs can (and increasingly do) delegate tasks for which they are structurally suboptimal (e.g., precise multi-step arithmetic or strict symbolic reasoning) to external tools or specialist modules. The model need only reliably detect the need, invoke a deterministic tool, incorporate the verified result, and resume fluent synthesis. This layered approach effectively mitigates the core weakness. It transforms an "idiot savant" into a highly capable orchestrator of precise computation.

Oh, what did they miss?

Oh did they figure out a mix of experts is better 😂 this is laughable, honestly, the whole AI world Is at the moment blind or something, reframing the same concepts over and over to get billions, re packing their old AI with new tricks, which FYI most Devs on git are better