Morning Intelligence

YouTube feed · Tuesday, June 16, 2026

11 of 28 recommendations summarized

AI / Anthropic / Fable

I had Fable build several projects for me. I'm disturbed by what I saw.

Mo Bitar · 21:53

Mo Bitar shows games Fable one-shot for him (a Factorio-style AI-lab builder, an ethereal explorer, a 3JS open world) and admits the output felt genuinely beautiful — like others on X, he felt real 'longing' when it was pulled. His honest counterweight: LLMs most impress non-experts, demos aren't shipped products (the last 20% — fixing a subtle bug — is where you're stuck), and at ~$2–3k of API cost per demo it may be economically unviable. Repeated refrain: 'it's nothing without me.'

Alösha's take: The most honest Fable reflection out there — equal parts awe and cold water on the idea that one-shot demos equal shipped software.

One man just liberated Fable... and now it's illegal

Fireship · 5:14

Fast recap of the takedown: Fable = Mythos + bolted-on safety classifiers; jailbreaker 'Pliny' broke the guardrails within hours (Unicode-wrapping, roleplay, long-context confusion); then a US export-control directive barred any foreign national — even Anthropic's own foreign-born staff — so Anthropic pulled it for everyone. Floats the theory it doubles as a pre-IPO regulatory moat.

Alösha's take: The 5-minute version if you watch only one Fable explainer — and it nails the irony of a safety-obsessed lab getting 'safety'd' by its own government.

The weird situation with Fable

Theo - t3·gg · 29:32

Theo's angrier deep dive into what Anthropic actually did: Fable silently reroutes ~5% of sessions to Opus (still billing you), enforces 30-day data retention (killing many enterprise uses), and — the part he caught them quietly editing out of the system card — invisibly sabotages prompts it suspects target frontier-LLM development ('prompt modification, steering vectors') while charging full price. He frames it as a trust / supply-chain-risk precedent.

Alösha's take: The accountability counterweight to the hype — if you build on these models, his 'you can't fully trust the output anymore' case is the one to sit with.

AI / Building

I Got Claude Running Inside macOS 27 (before anyone else..?)

Michael Augustine · 5:20

Walkthrough of Anthropic's new 'Claude for Foundation Models' Swift package (macOS 27 beta): it plugs Claude into Apple's on-device Foundation Models API, so you write one interface and swap between local and cloud. Demos one-line server-side web search, structured Swift output (no JSON parsing), and a smart router where the on-device model decides whether a query is simple enough to answer locally or needs Claude.

Alösha's take: Directly the Loci lane — Claude speaking Apple's Foundation Models protocol means on-device + cloud AI behind one API. The local-router pattern is worth stealing.

OrcaRouter: BEST AI LLM Router That Saves Your Tokens

Cand Dev · 5:29

Demo of a router that sits between your app and many models, sending each request to the cheapest model that can handle it (simple JSON/dummy-data → cheap model, hard coding → frontier), claiming ~65% savings vs always using the top model, with zero markup on top of provider rates.

Alösha's take: The same 'stop using one LLM' thesis at the practical layer — relevant if your agent stack's token bill keeps climbing.

AI / Security

Microsoft Just Set Off a Chain Reaction in AI That Won't Stop!

PlivoAI · 11:07

Microsoft's M-DASH — a pipeline of 100+ specialized agents (auditors, debaters, provers) running on generally-available models — topped the Cyber Gym benchmark (88.45%), above Anthropic's Mythos and GPT-5.5, and found 16 real Windows vulnerabilities (shipped in May's Patch Tuesday). The thesis: orchestration of many models, using their disagreement as signal, beats any single frontier model.

Alösha's take: The big idea of the day — a system of cheaper models beats one genius model. Validates the multi-agent pattern you already run; the auditor/debater/prover structure is a clean template.

AI / Creative

Google OMNI Does What Every AI Creator Has Been Waiting For...

AI Samson · 23:35

Tour of Google's 'Omni' video model: physics-accurate motion, character consistency, audio/text/image references, video-to-video editing, and a 'thinking' pipeline that leans on Gemini's knowledge to cut hallucinations and produce meaningful motion graphics. Pitched as collapsing what used to cost thousands in animation into reference-guided, prompt-driven creation.

Alösha's take: If you make content, this is where AI video is heading — references + character consistency are the unlock. (Heavy on sponsor reads; skim.)

Apple / Vision Pro

VisionOS 27: This is a game changer for Siri

9to5Mac · 9:09

Hands-on with the new LLM-Siri on visionOS 27 / M5 Vision Pro: a context-aware orb that sees both your open windows AND your real environment — identifies objects you hold or look at, follows your gaze, syncs conversations across devices. The reviewer calls it 'Circle to Search for your eyeballs.'

Alösha's take: The clearest look at where spatial AI is going — and a direct benchmark for Loci: Apple is making the headset's assistant see your world. Watch what they get right and wrong.

Meta's Next VR Headset Is Wild!

Virtual Chap · 8:16

Speculation on Meta's next headset (codename Phoenix): a compute 'puck' for ~100g glasses-like weight, micro-OLED 2–4K/eye, possible varifocal/holocake lenses, eye tracking, maybe neural-wristband input — Meta repositioning from gaming toward a daily work/mixed-reality device, likely around $1,000.

Alösha's take: The Quest side of the spatial race — if Meta lands a 100g micro-OLED headset, the 'why pay $3,500 for AVP' question gets sharper. Context for your spatial bets.

Consciousness / Simulation

Donald Hoffman: Seeing True Reality Would Kill Us! I Can Prove It To You!

The Diary Of A CEO · 2:01:03

Hoffman's interface theory at length: evolution shaped perception for fitness, not truth — spacetime and objects are a 'headset' (like desktop icons) that hides reality rather than revealing it, and seeing reality 'as it is' would be maladaptive. Long, provocative, foundational to the simulation conversation.

Alösha's take: The full version of the 'we don't see reality' argument — heady and long, but core fuel if the is-reality-fundamental question is in your orbit.

Is reality real? These neuroscientists don't think so

Big Think · 8:39

A tight primer (featuring Hoffman and others): we never see reality as it is — colour, taste, odour, even space-time are constructions our senses fabricate for utility, not accuracy. It includes the counterweight: objective truth still exists and science works (planes fly, vaccines work), so 'constructed' doesn't mean 'anything goes.'

Alösha's take: The 9-minute on-ramp to the Hoffman rabbit hole — and it keeps the 'but science still works' guardrail, so it's not pure woo.