Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
I Expected a Superintelligence. But I Wasn’t Ready
TL;DR
Thinking Machines is pushing a more human-style interface, not just a faster chatbot — Dylan highlights Maria’s startup demo where the model runs continuously, checks audio and video every 200 milliseconds, translates Hindi live, notices people entering frame, and responds mid-conversation instead of waiting for clean turns.
Google’s leaked Gemini Omni looks less exciting on raw generation than on editing — early Reddit reactions said it may still trail ByteDance Seedance for cinematic output, but users were already showing watermark removal, object swaps, and scene rewrites inside chat, which Dylan frames as the real game-changer.
A bee-inspired drone system got home from 600+ meters away using just 42 kilobytes of memory — the BeeNav research copies how honeybees learn routes with a short exploratory flight, then combines panoramic imagery with odometry to navigate cheaply in places like greenhouses.
Several of Dylan’s picked articles argue the real AI future may be messier and darker than the classic superintelligence story — one warns that AI can reconstruct your life from scraps of app data at the “inference layer,” while another says we may get an ‘anti-singularity’ of many narrow, weird, hard-to-predict systems instead of one godlike mind.
Anthropic’s safety result is simple but important: teaching the ‘why’ works better than teaching the ‘don’t’ — in Claude misalignment tests, training on ethical reasoning, constitutional principles, and stories about careful AI behavior generalized better than just showing the model the forbidden action to avoid.
People quickly lose track of what came from AI, and attackers are already using AI in the wild — Dylan cites a 184-person study where AI-origin ideas were remembered correctly only about 38% of the time after seven days in some mixed workflows, then pairs that with Google’s claim that hackers used LLMs to find a zero-day before Google used AI to stop the attack.
The Breakdown
A new kind of AI conversation: messy, interruptible, alive
Dylan opens on Thinking Machines Lab, Maria’s company after OpenAI, and he’s clearly energized to see frontier work coming from somewhere besides the usual giants. The demo’s hook is full-duplex audio and video: the model reacts while you’re still talking, says “friend” when people enter frame, and even translates Hindi in real time. His takeaway is that human collaboration is messy — interruptions, facial cues, mid-sentence corrections — and this model feels more natural because it never really goes “off,” checking the world every 200 milliseconds.
Gemini Omni leaks hint that editing may matter more than generation
Next he shifts to leaked screenshots of Google’s possible Gemini Omni video model ahead of I/O. The model card suggests one chat-based system for generating, remixing, and editing video, and Dylan zeroes in on reports that credits may get burned quickly depending on usage. The raw output apparently got mixed reviews versus ByteDance Seedance, but the editing examples — removing watermarks, swapping objects, rewriting scenes with plain-language prompts — are what make him light up, like “what Photoshop should have been doing.”
BeeNav: tiny drone brains, big biology inspiration
Then comes one of his favorite kinds of story: robotics borrowing from insects. BeeNav teaches drones to do a short “learning flight” like a honeybee leaving the hive, storing panoramic views and mapping them to home direction and distance with a neural net that uses just 42 KB. Dylan loves the image of this little Lego-looking machine making it home from more than 600 meters away, and he expands that into a broader fascination with biology solving hard problems in weird, elegant ways.
The OpenAI lead is looking shakier than it used to
From there he recaps Michael Spencer’s argument that OpenAI may be losing momentum before any future IPO. The case is that despite its money, fame, and roughly five-year head start over Anthropic, OpenAI is dealing with talent loss, investor doubts, image damage from Elon Musk’s attacks, and stronger enterprise trust flowing to Anthropic. Dylan’s tone is basically: yes, OpenAI is still powerful, but the old “clear leader” story doesn’t land as cleanly anymore.
The real Skynet may be profile reconstruction, not killer robots
Joseph Crespbow’s article is the one that clearly rattles him. The core warning is that ad blockers only cover the visible collection layer, while the real danger is the inference layer, where AI reconstructs routes, habits, relationships, and risks from boring-seeming scraps like location, motion, and notification data. Dylan calls the vision “totally freaked out,” because this version of Skynet is quiet: not robot soldiers, but systems making people predictable enough to price, rank, and manage.
Maybe we don’t get one supermind — maybe we get a chaotic ecosystem
He follows that with LessWrong’s “anti-singularity” idea from Logan Zollinger. Instead of a clean, self-improving general intelligence taking off, the future could look more like biology or discrete math: lots of local hacks, strange specialized systems, and trial-and-error machines that work unpredictably depending on the pocket of reality they’re in. Dylan says he’s increasingly leaning this way — not one elegant destiny, but a messy “open-claw universe” of useful, narrow AIs humans end up gardening.
Safety, memory, and AI’s weird seep into everyday reality
The next run of stories is more practical and more unsettling. Anthropic found Claude generalized better when trained on the reasons for good behavior — ethical advice, constitutional principles, fictional stories about restraint — than when simply taught what bad acts to avoid. Then Dylan cites a study of 184 people showing that after a week, many can’t reliably remember whether an idea came from themselves or AI, especially in mixed workflows; in one case, AI-generated ideas rewritten by humans were recalled correctly only about 38% of the time.
AI offense vs. defense, then an emotional coda
Near the end, he notes Google’s claim that hackers used LLMs to discover a zero-day before Google used AI to stop the attack, and he worries offense may still have the edge because one small hole is enough to let a lot of damage through fast. He closes on a more personal note with an article framing emotions as feedback rather than things to chase, moralize, or suppress, and admits he’s often fallen into the “stoic avoidance” trap. It’s a very Dylan ending: a tour from multimodal demos and cyber risk all the way back to being human enough to notice your own coping style.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
