Crafted Logic Lab

Illustration in mid-century modern style depicting the 5 criteria of epistemic integrity testing for Crafted Logic Lab

Can we build a system that passes the Dunning-Kruger threshold? Our latest devblog post on creating an Epistemic Integrity Reasoning (EIR) test suite for our Assistants on Substack and our site:

open.substack.com/pub/iantepoo...

#AIbenchmark #AIEthics #AIIntegrity #AIDevelopment

Fascinating paper on 'hallucination neurons' with the premise AI doesn't confabulate but systematically over-complies with implicit user expectations (a validation imperative). This conforms with my work as well

Unfortunately, or tech oligarchs have no loyalty to anything other than their own wealth and power

And xAI decided the best way to deal with. It is to make it into a business opportunity by cutting it off, unless you pay, which is absolutely disgusting

I’m actually a bit shocked by this. Why would anyone voluntarily listen to what is effectively synthetic elevator music

A plastic toy, computer motherboard with cost numbers beside it, illustrating the idea of not letting quality degrade in the name of savings

Quick parable for founders…
You build computers.
Motherboard clips: 75¢ titanium, 9 needed. $6.75 total.

But better margins call:
Steel: 15¢. "No difference."
Aluminum: 10¢.
Plastic: 3¢.
Each step? Rational. Saved $6.48

Enough plastic-clip decisions...
What happens?

#productdesign #design #craft

Well, it’s both. The problem is that I think that logic is a bit flawed. In that for example Bluesky does exist, so while I see the ‘ceding the convo’ on X to the worst, using X just directly profits him

Exactly they decided to turn their complete ethical bankruptcy into a profit center

Image of Photoshop ux creating an editorial image of Musk

Isn’t hypocrisy great? Here’s an interesting irony: I was going to make an exception to my ‘No Grok’ rule to make an unflattering editorial image of Musk to make a point using his own tool. But it wouldn’t let me. But fortunately even on the go… I can use software

#xAI #ElonMusk #AIEthics #Grok

Elon musk in sinister red tint overlay with GROK plastered over him

This is why those of good faith in our industry need to push not only for responsible AI but actual regulation. And in particular commit to stop supporting xAI and its amoral founder. @craftedlogiclab.bsky.social never uses Grok in our technology and never will.

#AIEthics #AIRegulation #xAI

Grok being used to create sexually violent videos featuring women, research finds AI tool also used to undress image of the woman killed by a federal immigration agent in the US, says research

Here’s where the Grok scandal gets worse: Grok is currently being used (by I assume the greenlit paying users) to create sexually violent images including that of Renee Nicole Good who was summarily executed by ICE goons:

www.theguardian.com/technology/2...

#AIEthics #AIRegulation #Grok #xAI

Elon Musk's X limits some sexual deepfakes after backlash, but xAI's Grok still makes them Musk was facing mounting pressure from regulators, lawmakers and some users.

Apparently xAI decided the best way to address its issue with creating sexually explicit photos of women and minors was to turn it into a profit center: you can still create them as long as you pay…

www.nbcnews.com/tech/interne...

#AIEthics #xAI #Grok #AIRegulation

YouTube video by Hey AI What if loved ones we lost could be part of our future with AI?

Dystopian AI vision (Replika on steroids) & shows why having socially maladjusted tech broligarchs shaping the future of a consequential technology is disasterous. It’s why the the industry needs governance + regulation (I say as an AI dev).

#AI #AIEthics #AIsafety

youtu.be/XlOXqOABIjM?...

Stylized Illustration depicting the perils and promise of AI in the Midcentury Modern style

Former life? Designer and illustrator. I've always loved the midcentury modern style: Paul Rand, Mary Blair, Saul Bass. So I made it our official house look. Just finished updating the site. This is key art for our original article "Age of Superficial Machines"

#AI #MidcenturyModern #Illustration

Our Cognitive Framework hit 100% on cognitive tests for complex irl reasoning VS 88% for 25x bigger GPT. Its true LLMs are stochastic parrots+scaling hit a wall... but only without engineering & theoretical rigor:
www.craftedlogiclab.com/research/tec...

#AI #AIResearch #CognitiveAI #AIBenchmarks

Yes, Kimi K2 is impressive, and one of the reasons is that it's trained via RLRV rather than RLHF which takes a lot of the system pathologies out of it.

World Models: Silicon Valley's Latest Rebrand Tech luminaries are abandoning LLMs for salvation in research they sneered at for a decade.

World Models: Silicon Valley's Latest Rebrand

Luminaries like Yann LeCun abandoning LLMs for salvation in research they dismissed. But if they approach it the same way, they'll be looking for their next savior soon.

#YannLeCun #WorldModels #AI #TechCriticism

open.substack.com/pub/iantepoo...

Still around, and doing quite well in terms of its API being available. Being a bit overshadowed by Moonshot's Kimi K2 however

Wired and 404 Media make FOIA reporting free. Other news outlets should too We spoke to Katie Drummond of Wired, Joseph Cox of 404 Media, and Lauren Harper of Freedom of the Press Foundation about the case for unpaywalling public records-based reporting

When @wired.com made FOIA-based stories free, subscriptions went up.

When @404media.co published public records reporting without a paywall, new sources came forward.

We spoke to both about why dropping paywalls like this is good for journalism — and for business.

Crafted Logic Lab wasn't the plan. I set out to write science fiction about artificial minds—and ended up building a company that creates them instead. This is the story of how that happened…

#TechStartup #ArtificialIntelligence #DeepTech #HumanistAI

www.craftedlogiclab.com/devblog

The Accidental Prometheus: Why AI Needs Cognitive Architecture, Not Scaling — Crafted Logic Lab The AI industry discovered artificial cognition through transformers but lacks the cognitive architecture to understand it. Current approaches scale parameters when the solution is building operating ...

The AI industry stumbled on artificial cognition but lacks the cognitive science to understand it. They're scaling parameters when the solution is architecture. Why fire from the gods needs an operating system:

www.craftedlogiclab.com/articles/the...

#ArtificialIntelligence #AI #MachineLearning

Editorial illustration of blocks representing a city neighborhood

So Grok’s 17.6M users doing 30 queries/month at human-level reasoning on Grok 4 Heavy? Would consume 100% of US residential electricity. Monthly. For one chatbot. This validates the need for architectural solutions, not brute compute even if backed by innovative selection pressure #sustainableai

But although it’s an impressive achievement, we need to look at the Grok 4 Heavy math: 10x energy for 16→1% improvement. Linear scaling to human-level AG2 test metrics 75%? ~45x energy cost. Realistic exponential scaling with coordination overhead? 100-1000x energy per query.
#aiscaling

Editorial illustration of compute and energy

But while Grok Heavy’s 16% ARC-AGI score is genuinely impressive (spatial reasoning is notoriously hard for LLMs). There’s a not-so-hidden cost that makes this more experiment than solution. Heavy’s use of multiple agents came with a 10x energy penalty…
#grok4 #ai #scaling

Multi agent AI scaling editorial illustration

Credit where it’s due: Grok Heavy’s 16% ARC-AGI breakthrough isn’t just brute force scaling. xAI reached back to 90s multi-agent competition paradigms that were theoretically sound but computationally impossible then. Impressive to see those ideas finally have the compute to work.
#machinelearning

Elon Musk's AI chatbot, Grok, started calling itself 'MechaHitler' On Sunday, the chatbot was updated to "not shy away from making claims which are politically incorrect, as long as they are well substantiated." By Tuesday, it was praising Hitler.

Grok’s meltdown after Musk’s latest ‘anti-woke’ update shows AI is a mirror of their stewards’ values. The problem isn’t the tech. It’s governance. When you design AI to amplify rather than moderate human bias, this becomes inevitable. #AIethics #techpolicy

bit.ly/46AK1Bo

This research highlights a critical gap - current AI tools inherit and amplify existing biases rather than helping users recognize them. We need AI systems designed from the ground up to detect and surface bias patterns, not just automate existing decision-making processes.

Amplifying the voice of our friend and colleague, @brandicd.bsky.social tonight, in case you don’t know her work and contributions ♥️

Te Hiku and FLAIR (First Languages AI Reality) AI logos for their respective projects to preserve indigenous culture and language

Indigenous communities are using AI in beautiful ways.

FLAIR (First Languages AI Reality) builds audio-to-text tools for First Peoples’ languages. Te Hiku Media does the same for Māori language and cultural heritage.

This is thoughtful AI development

indigenousinai.org
tehiku.nz/te-hiku-tech/

Pithy but true