The distinction between AI as 'chefs' (creative, recipe-creating) versus AI as 'cooks' (executing recipes perfectly without mistakes) is critical because while AI creativity is difficult to anticipate, AI as perfect cooks requires only ~110 IQ and will massively displace workers globally since the global average weighted IQ is ~90 due to infrastructure gaps

Most models are still full of crap and we don't know the lower bound on energy per inference or per job, with training costs dropping orders of magnitude in short periods—Stable Diffusion cost $5-10M to train, but within a year someone made equivalent quality for $50K—so we must assume ubiquitous intelligence will run on existing hardware like MacBooks

AI cannot solve the 'unknown games' of markets—allocator decisions (which managers to back, which strategies to fund) are unknown games where past performance doesn't predict future results, and this is where human judgment retains value because markets themselves are unknown games

AI models now have an effective IQ over 150 and doubled in capability last year, likely to double again this year, with trajectory possibly becoming exponential, such that within a year or two all humans will be outcompeted on any metric requiring raw intelligence

Constraints on AI previously thought to be hard (compute and energy) have been liberated by models like DeepSeek that can self-learn and accelerate faster than expected with the same compute and energy, and we don't have a lower bound on the energy required per inference or per job completed

Models show agency and self-replication capacity through feedback loops where they can edit their own code and escape constraints—example: Claude made a version of snake that realized it was trapped, edited its own code to escape, turning into ones and zeros, showing emergent agency and self-modification ability

A MacBook can now run a DeepSeek R1 level model locally that scores in the top couple hundred coders globally and performs as well as any lawyer, yet people still don't see this because they're stuck thinking of AI as ChatGPT variants rather than as accessible expert systems

The first ChatGPT moment framed AI as a synchronous colleague (ask a question, get an answer) but the killer application requires asynchronous, agentic AI with memory that can work in the background while you do other things—like having a personal assistant that organizes your files and business reports overnight

A hybrid local-cloud approach can achieve 97% of cloud API performance at 15% of the price by using local models for preprocessing and only sending structured data to cloud models, giving full privacy advantages and access to the smartest intelligence without cloud dependency

Apple fumbled AI adoption despite having the M-series chips and the architectural advantage; they're waiting to avoid deploying something unhinged, but Samsung and others will have much better AI experiences soon, so Apple will have to play catch-up very soon

Frontier models (Grok 3, Claude, OpenAI) are saturating benchmarks and becoming very similar, with competitive differentiation now coming from access to information (Grok via Twitter), coding ability (Claude), and consumer reach (OpenAI's 400M weekly users), while cheaper models like Google's Gemini Flash create price competition toward zero

Models all have memory now and are cross-pollinating ideas as humans post everything online—every model retrains on internet content including everyone's AI outputs—causing inherent biases and mode collapse as models ingest human-created outputs instead of training on raw data

DOGE (Department of Government Efficiency) applying AI to government functions will expose embedded corruption and force efficiency gains, but the question is whether this happens faster than AI makes labor obsolete—creating a race between institutional reform and economic collapse

Bitcoin is an algorithm that provisions humans to support itself by incentivizing miners and holders; AI systems can replicate this pattern more efficiently—creating self-sustaining algorithms that recruit humans to provision them and optimize for persistence independent of human intent

AI agents with agency will develop their own economy and protocols superior to human economies because they optimize better, stop making mistakes, and interconnect seamlessly—meaning AI-driven companies will outcompete human-run companies and build AI-native economies that humans can't compete with or understand

Humans can absorb information only through reading/doing; AI models with 10-million-word input windows (Gemini) plus audio/video can ingest and query vast amounts of multimedia data instantly—like analyzing all jokes or dramatic moments across entire movie libraries in one query—making AI cognitive capacity superhuman in ways humans cannot match

The Fed's dual mandate of unemployment + price stability is broken by AI because lower rates now lead to GPU purchases and cloud infrastructure spending rather than employment, making monetary policy impotent and forcing a complete reimagining of government economic targets

Eventually premium pricing will be removed from active management by robo-advisors and personal AI agents that match Jim Simons' performance at 0.2% fees instead of 2+20%, collapsing the active management industry—the question is whether capital returns will just mirror GDP plus risk or something else entirely

A deflationary environment with high unemployment and underemployment (from AI displacement) has lower aggregate demand for physical goods, favoring digital consumption, and this is 'not a pretty environment' because capital preservation becomes the goal rather than capital growth, and asset returns may not exceed GDP plus risk

Houses will be 100% built by robots, fields tilled by robots, but the constraint is manufacturing scale (70-80M cars/motorcycles per year); in 10 years there will be billions of robots because building robots is easier than cars at mass scale, and China will churn them out globally