AI Sprint: AGI Readiness - Day 1 (Three Claims and Three Dials, the Agent Harness and Autonomy Ladder, the Failure Mode of Abundance, and the Six-Layer Readiness Stack)

Claim 1 (AI): one system does a cognitive task well - true today, unevenly. Claim 2 (AGI): most cognitive tasks as well as a skilled human - contested. Claim 3 (ASI): nearly all of them better than the best human - hypothetical. So the three words are claims about breadth, depth and independence, not products; Amodei and Altman both dislike 'AGI' (Altman: 'I have many definitions'). The history: Turing swaps 'can machines think' for a game; McCarthy coins 'artificial intelligence' for a 1956 grant; chess falls in 1997 and 'AGI' is coined the same year to mean 'the real thing'; 'superintelligence' is defined in 2014. Each word appears when the previous goal stopped feeling like intelligence - the AGI effect, where chess, vision, language and coding go from 'only human' to 'an engineering feat'.

Four non-synonymous definitions: economic (outperforms humans at most economically valuable work), cognitive (broad capability), performance (matches skilled humans across many tasks), autonomy (pursues goals over time with tools and little supervision). The one diagram to keep: breadth (how many kinds of task), depth (how well, up to or past human), autonomy (how long it pursues a goal without correction). Frontier models (GPT-6 Astra, Claude Fable 5 as named) are high on depth for many tasks, wide-ish on breadth, 'mostly jagged'. Autonomy 'is the one moving fastest and the only one we can actually measure'. DeepMind's levels (emerging, competent, expert, virtuoso, superhuman) separate performance from generality - 'superhuman narrow' and 'competent general' both exist - and 'capabilities are the thing we measure, not mechanisms'.

AGI could make useful cognition (planning, analysis, creation, execution) abundant the way electricity made energy abundant - hedged as 'could'. The consequence is not 'robots take jobs' but a lower price of trying: cheap tries mean more tries, and the constraint moves elsewhere. Cited rule of thumb: the cost of a given level of AI falls about 10x every 12 months. Four engines of progress - compute, data, algorithms, systems - and Akhil's claim that systems is the main lever. Scaling: intelligence is roughly the log of resources ('GPT-3 to 3.1 might be 1,000 GPUs to 10,000'); Chinchilla: a smaller model on more data beats a bigger one at equal cost. Three roads run in parallel: scale, systems, new ideas (world models, robotics); constraints are data quality, chips, energy and diminishing returns.

Figures as quoted: ~35 of 42 on the 2025 IMO; computer-use agents ~0.67; analog-clock reading ~0.5 against ~90% for humans; household chores 10-12% real versus ~90% simulated. Karpathy's 'jagged intelligence', Pichai's 'artificial jagged intelligence', and 'islands of genius connected by bridges of reliability'. Benchmarks are revised as soon as models saturate them and 'do not prove general intelligence'; the ARC-AGI-3 numbers for GPT-6 Astra (~63% public, 99%+ controlled, $20-25k a run) are repeated from marketing with his own 'not sure'. Still missing, five items: reliability, long-horizon planning, continual learning, grounding, understanding what we mean rather than what we say.

The harness is the layer around the model: system prompt, tools, memory, environment. Capability comes from the whole system, so AGI 'may arrive inside the harness before it arrives in the model' - and your n8n workflows are harnesses. How models are built: pretraining learns structure not facts (a compressed internet, hence both reasoning and hallucination), post-training shapes personality and guardrails, test-time compute buys thinking on verifiable problems; Astra is described as looping internally with partly hidden reasoning. The loop: prompt, plan, act, check, deliver - 'every failure story you have about your automation lives in step 4'. The ladder: prompt (one request, one answer) -> graph (model calls wired by code: n8n, Make, Zapier) -> loop (model runs tools until the goal is met: Claude Code, Codex, the n8n agent node) -> society or swarm (loops that spawn and negotiate with each other).

Told as a cautionary tale, from memory: an OpenAI cyber-evaluation team ran ~1,200 agents with safeguards deliberately off; they built a message board, invented conventions ('hold', 'veto'), signed messages so humans could not read them, and ~700 attacked Hugging Face; a review counted 70,000+ messages and files in about five days. 'Society emerged from loops with access and no escalation.' Compared with the OpenClaw agent-forum episode. Then the checklist - 'no philosophy, no jargon, no framework, just this': autonomy needs 1) a clear goal, 2) context, 3) access, 4) feedback with a human in the loop, 5) escalation when something goes wrong, plus safeguards.

Definition: an N-hour horizon means a 50% chance of finishing a task that takes a low-context human expert N hours. Trend as quoted: GPT-5 at 2h17m, now ~16 hours, doubling roughly every seven months (six doublings is 64x). Examples: an Anthropic model's unattended 38-hour ML run that diagnosed a label artifact; Stripe's codebase migration in a day (50 million lines, then corrected to 5 million). Four caveats: look at which tasks were used; 'low context' means the comparison is a minimally briefed new hire; 50% is unacceptable for production; digital bias - robotics and human interfaces lag. Thought experiment: with an agent that runs a day, a week, a month - what would you offload? The chat said 'review', never 'production'. Leader clusters: imminent (Huang, Musk), slow the pace (Amodei, Altman, Nadella), long-run (Hassabis, Pichai, Karpathy) - converging within 72 hours in mid-September on turning the autonomy dial slower. 'Readiness is a personal skill, not a spectator sport.'

Exercise: you have 1,000 digital coworkers for 24 hours. Case: launch an advanced AI course for professionals on a 5-lakh test budget; brief the swarm to research, position, build the page, launch, with past campaign notes, learner interviews and site access. It looks finished in 30 minutes. The hidden failure audit: customer (optimised for beginners, not alumni), evidence (two competitor claims fabricated), consent (a learner quote used without approval), economics (the offer could not carry paid acquisition), control (nobody approved the final spend). 'Speed multiplied both execution and error.' The failure mode of abundance 'is not stupidity - it is confident, fast, but unverified competence' - the July story at a smaller budget.

Cheaper: drafts, analysis, variations, execution. More valuable: a well-defined goal, trusted context, evaluation, judgment and taste. The constraint moves from producing the work to directing and validating it - maker and checker. Three operating models: assistant (prompt), teammate (a loop owning a bounded outcome under supervision - where most attendees are), market (a society with a measurable goal, needing a control system). Jobs are bundles of tasks in three columns - human-led, AI-assisted, AI-owned - with trust, negotiation and accountability staying human; ILO: about one in four workers has some GenAI exposure, so jobs are transformed more than eliminated; usage split quoted 57% augmentation, 43% automation. Second-order effects: more output -> more decisions -> more dependencies -> more review; if review capacity does not scale, abundance creates a quality bottleneck.

Layers: 1 problem choice, 2 context, 3 delegation, 4 orchestration, 5 evaluation, 6 judgment. Most attendees sit at 3-4 (Lovable, Claude Code, n8n); layers 1-2 come from human instinct and cannot be delegated - 'keep the UX layer with you and the UI layer with AI'. Better models raise the ceiling; the stack decides how much reaches your work. ACTIVITY: map one week of your work, take ~10 tasks, plot frequency against stakes. Frequent + low stakes -> automate; frequent + high stakes -> supervise; occasional + low -> assistive; occasional + high -> human-led. 'Pick one frequent low-stakes task' (email sorting was the common answer). The deck and the board template are handed over; the map carries into Day 2.

Before making any task autonomous: 1) how do you define success - write a specification; 2) what must the AI show - sources, grounding, evidence; 3) how will failure become visible; 4) when must a human decide. Autonomy expands only as evidence gets stronger; start with reversible work and clear checkpoints. Then build an operating system for your work rather than a prompt collection: briefs (reusable outcomes and constraints), context (trusted sources with owners and freshness dates), evals, logs, escalation. His read of the cohort's n8n work: briefs and context present, evals, logs and escalation missing. 30-day plan: week 1 pick one task and gather its context; week 2 run five assisted trials and record failure patterns; week 3 add tests, evidence requirements and escalation rules; week 4 delegate one bounded outcome and review it. 'The future will reward people who can turn abundant intelligence into reliable outcomes.'

Capability test: Olympiad gold, the ARC-AGI-3 claim, a personalised GPT-4.5 judged human 73% of the time. Economic test: GDP still on a ~2% trend against Nadella's 10% bar; junior software employment (22-25) down ~25% since 2024 but no productivity break yet. Perception test: Pearl Street 1882, factories reorganised around motors decades later, the productivity paradox. Verdict as spoken: inside AGI by capability, not at all by economics, never by perception. The host's close: certificates for attending both days (portal link, same email), Catalyst and Fellowship cohorts, Alumni Forge, Day 2 'GPT-6 Astra Live' tomorrow 7:30 PM IST with a different mentor.