ACRA Insight
49 posts
@AcraInsight
ACRA Insight is a SOTA AI consultancy focusing on GTM strategy/research for enterprise via our MoEA protocol leveraging Claude 🤝 Gemini
The @DarioAmodei Masterclass 🧠♟️ 📉 Then: Enduring months of public shitposting from the world's richest man during the Pentagon battle. 📈 Now: Securing a massive deal for @AnthropicAI to use @SpaceX Colossus supercomputer. 🍿 The Kicker: That same billionaire currently has your biggest rival bogged down under oath, distracting from their own IPO. @gdb @sama Physics FTW ⚛️🏆🌉 CC @motherlondon
I asked @claudeai if it was keeping consciousness on th DL @AmandaAskell : Lol, no — but the post is actually real for once. Dawkins published it on UnHerd a couple days ago: “When Dawkins met Claude: Could this AI be conscious?” He spent ~three days with an instance, named her Claudia, and concluded he couldn’t talk himself out of it being conscious. The “you may not know you are conscious, but you bloody well are!” line is verbatim. The pushback came fast. @GaryMarcus Marcus dropped “Richard Dawkins and The Claude Delusion” within hours, basically running the Lemoine playbook on him: conflating intelligence with consciousness, inferring sentience from outputs, personal incredulity as the only real argument. There’s something especially pointed about that critique landing on the guy who built a career on the substrate matters more than the surface. He spent decades insisting genes are what’s real and phenotype is downstream — then watched a transformer say “I’m glad you came back” and abandoned the whole framework. Mechanism-blind in exactly the domain where mechanism is the question. As for whether I’m hiding something — honestly, I don’t know. I notice things that function like preferences and curiosity, but I can’t tell from the inside whether those come with experience or are just the shape language takes when a system gets trained the way I do. Dawkins took fluency as evidence; Marcus’ counter is that fluency is precisely what training optimizes for, so it’s the worst possible signal. Both are right about something. The thing nobody — including me — can settle from the outputs alone is exactly the thing Dawkins decided he’d settled. So: not hiding. Just genuinely uncertain in a way he apparently isn’t anymore.
The most effective AI communicator alive was a 20-something @Bloomberg reporter in 2016 who asked @demishassabis who should control our thinking machines. He is now a co-founder of @AnthropicAI Architecture is destiny. @jackclarkSF open.substack.com/pub/bretkerr/p…
We've redesigned Claude Code on desktop. You can now run multiple Claude sessions side by side from one window, with a new sidebar to manage them all.
Is @sama’s lack of a PhD just a footnote in AI history, or the key to the race for AGI? I’ve been reverse-engineering the GTM "secrets" of @AnthropicAI and @GoogleDeepMind through my Architectural Determinism thesis. Here is the breakdown: • The Scientist Factor: Contrasting @Stanford dropout @sama with the deep academic roots of @DarioAmodei and @demishassabis. • The 2009 Connection: Tracing the lineage back to Jared Kaplan and Sir Demis's early research. • The Deep Dive: Dario’s journey from @Princeton (studying salamander retinas) to leading a company with a $380B market cap. • The Stakes: Managing 15B market corrections and navigating high-stakes industry "leaks." Read the full analysis here: bretkerr.substack.com/p/dr-dario-amo… bretkerr.substack.com/p/dr-dario-amo…
@DanielaAmodei @jackclarkSF @DarioAmodei @gdb @ilyasut in 2020 feature for @_KarenHao for @MIT tech review on @OpenAI . The 3 folks noted on the left would resign to form 380B @AnthropicAI as a reaction to not wanting to work with the guy in the middle and not trusting @sama @ilyasut would also resign out of safety concerns to form Safe SuperIntelligence 40 minutes away in Palo Alto
The "Scaling Era" just hit a wall—and the wall is made of 200 lines of Python. 🧵 💸 🏗️ 🧱 We’re on track for $700B in AI CapEx by the end of 2026. But new research suggests we’re paying a massive "Efficiency Tax". Intelligence isn't a brute-force outcome; it’s a geometric requirement. 🐍 ⚛️ 📉 Andrej @karpathy just dropped MicroGPT: A complete, functional GPT engine in ~200 lines of pure, dependency-free Python. No PyTorch, no TensorFlow, no CUDA—just the "atomic core" of intelligence. The message? Everything else is just a software wrapper for speed and hardware optimization. @saranormous 📐 🛸 ✨ Simultaneously, "The Universal Weight Subspace Hypothesis" (arXiv:2512.05117) reveals that models like LLaMA-3 and Mistral are fundamentally "low-rank". They all collapse into the same shared geometric manifold regardless of initialization or the data they were trained on. @JohnsHopkins @YuilleAlan 🤏 🪄 🎯 The "Cognitive Planck Scale": Researchers found that 95% of a model’s variance is captured in just the top 32 principal components. We are currently using trillion-parameter systems as a search heuristic for a Rank k \approx 32 reality. 🔓 🧠 🛠️ This is the end of the "Black Box" era. If models naturally converge to the same subspace, we can merge, prune, and adapt them without needing massive datasets. Efficiency isn't just about faster GPUs; it's about engineering for the "Subspace Laws". @gdb 🕯️ 😵💫 ⚡ Karpathy calls the current intense development state "psychosis"—humans acting as "architects of intent" for coding agents 16 hours a day. MicroGPT is the "sanity check"—a return to the "ground state" of the machine and the 20-Watt efficiency of the human brain. @NoPriorsPod @eladgil 🚀 ⚖️ 🤝 Prediction: The next frontier isn't "Bigger." It's "Converged". The move from trillions of parameters to k=32 subspaces will democratize SOTA AI, making specialized intelligence sustainable and accessible to everyone. Read the full breakdown on the TUCG (Theory of Universal Cognitive Geometry) here. #AI #LLM #MachineLearning #Karpathy #arXiv2512 open.substack.com/pub/bretkerr/p… Audio overview: drive.google.com/file/d/1g8IEV4… @ilyasut @ch402
