my aha moment:
i've opened whatsapp and enabled siri and asked check what this person wants and answer (it was my calendar availability).
this deep os/hardware integration is the moat.
and @Apple will absorb a lot of llm calls on-device because in 80% of cases this model intelligence will be enough.
@tryramp need to make AI gateway similar to @LiteLLM our even buy them. Budgets, policies, etc on tokens spend. Later routing. Better visibility compared to just invoices.
I’m on @emirates flight to London which means 7 hours of vibecoing plugin for @_HermesAgent for open source personal finance assistant.
I suggest to reposition codex from “do almost anything” to “do anything anywhere anytime”.
I decided to re-read “Situational Awareness” by @leopoldasch, and here’s a comparison of what he predicted in 2024 vs. what actually happened by 2026.
Scaling will keep working: Right.
GPT-5.5, Claude Opus 4.7, Gemini 3, DeepSeek V4/R1, Kimi K2.6 all kept pushing frontier capability forward.
Agents will move from chatbots to coworkers: Right.
AI systems now browse, code, research, use tools, operate computers, and execute long-running workflows semi-autonomously. Humans increasingly supervise rather than manually execute.
GPUs/compute become the core bottleneck and strategic asset: Very right.
AI became constrained by GPUs, power, cooling, energy, and data centers. NVIDIA effectively became core global AI infrastructure. Stargate and sovereign AI factory projects validated the “compute is destiny” thesis.
China/open models become major competitors: Right.
DeepSeek, Qwen, Kimi, GLM, and others dramatically closed the gap on reasoning and coding, especially on price/performance.
Cyber/security becomes a major AI capability risk: Right.
Frontier cyber models are now increasingly restricted and treated as national-security-sensitive capabilities. Claude Mythos and GPT-5.5-Cyber are strong examples of this shift.
AGI by ~2027 is plausible: Still unclear.
Feels significantly more plausible in 2026 than it did in 2024, but still not proven.
Intelligence explosion after AI automates AI research: Not validated yet.
AI heavily accelerates coding, research, and experimentation, but there is still no evidence of a recursive self-improvement runaway loop.
Institutions are unprepared: Right.
Capabilities moved faster than regulation, enterprise governance, security readiness, and societal adaptation.
A preview for Pro users: a new personal finance experience in ChatGPT.
Pro users in the U.S. can securely connect financial accounts, see where their money is going, and ask questions based on the information they choose to connect.
Your full financial picture, now in ChatGPT.
A preview for Pro users: a new personal finance experience in ChatGPT.
Pro users in the U.S. can securely connect financial accounts, see where their money is going, and ask questions based on the information they choose to connect.
Your full financial picture, now in ChatGPT.
@masonwarner I think bar is getting higher this days everyone has access to the same tools. We all need to rewatch “Mad Men” to make next “I’d like to buy the world a coke”. Creativity is the moat.
idea of "ai agents work while you sleep" is just a propaganda -> demand shaping strategy. GPUs at night heavily underutilized.
i think @OpenAI@AnthropicAI need some sort of Uber Surge Pricing. Social equilibrium as objective function ;-)
i build surge few times i see potential efficiency gain and hate
@WHOOP so instead of having tons of MCPs tools just expose /ask endpoint as integration and let 2 agents just talk to each other.
@grok what do you think?
Look, I did integration between Hermes Agent and @WHOOP API. Hermes is hitting /activity endpoint and doing own interpretation of sleep quality from response instead Hermes agent should hit /ask endpoint for Whoop AI Coach with question "how was Andrey's sleep" and WHOOP assistant will come up with expert level answer.
This is universal communication protocol -> natural language. human<>human, human<>ai, ai<>ai.
@MarkOfDaDon@OpenAI I do have Hermes agent running but use-case is more like reminders on steroid and it knows based on WHOOP when I wake up. Main issue is compliance, security, and friction over collecting Company Brain in a cloud and gateway to it with a phone.