AI: From the First Algorithm to Superintelligence — Complete Stream Outline

Live stream · February 22, 2026 · 63 min · Watch on YouTube

Overview

This is a 90-minute rebuild of a February 2026 live stream about artificial intelligence, built from the transcript and expanded with the frontier material that was promised but cut off. The stream covers 90 years of AI history — from Alan Turing's 1936 foundational paper through today's frontier models — and lands on what Miami-Dade County must do to prepare.

Each segment below has a target time, the opening beat, the talking points in order, and a hand-off into the next segment. Read the talking points as a reference guide, not a script. The delivery is conversational. Keep the energy direct and the vocabulary precise — this is real technical work explained clearly.

How to Use This Document

Two things to know if you are reading this instead of watching the stream:

The first 55 minutes retrace the February stream at a tighter pace; the last 35 minutes are new material covering AGI, superintelligence, alignment, jobs, and Miami-Dade readiness.
Two production improvements for a re-run: Front-load housekeeping (social media posting, co-host invites) into the first 60 seconds so it does not interrupt the Turing story. Put a hard clock on Segments 2–8 so you reach the frontier material with 35 minutes to spare.

Run of Show (90 Minutes)

First 55 minutes: Original stream re-paced

Last 35 minutes: New material

0:00 – Segment 1: Cold open: the 1936 library (4 min)
0:04 – Segment 2: Foundations: Church, Turing, halting problem (7 min)
0:11 – Segment 3: Dartmouth 1956, expert systems, AI winters (6 min)
0:17 – Segment 4: Backprop to ImageNet: the machine learns to see (7 min)
0:24 – Segment 5: The transformer revolution (8 min)
0:32 – Segment 6: Scaling, emergence, RLHF, ChatGPT (7 min)
0:39 – Segment 7: Multimodal, AlphaFold, agents, benchmarks (8 min)
0:47 – Segment 8: Inference and precision: why numbers cost money (8 min)
0:55 – Segment 9: AGI and superintelligence (9 min) — NEW
1:04 – Segment 10: The alignment problem (9 min) — NEW
1:13 – Segment 11: Jobs and Miami-Dade readiness (10 min) — NEW
1:23 – Segment 12: Close and Q&A (7 min) — NEW

Checkpoint: If you are not into Segment 9 by 1:00, skip the precision worked examples in Segment 8 and jump to the frontier hour. That is the material that matters most for Miami-Dade.

Segment 1 — Cold Open: The 1936 Library (0:00–0:04)

4 min

The opening beat: A thought experiment set in a university library in 1936, across the table from a 24-year-old Cambridge graduate student writing about a hypothetical machine on a tape.

Talking Points

"Imagine you are sitting in a university library in 1936." Across the table is a 24-year-old Cambridge graduate student. Not famous. Published nothing. He is writing about a hypothetical machine that reads symbols from an infinite tape, moves left and right, writes new symbols, and halts. His name is Alan Turing. That paper, "On Computable Numbers, with an Application to the Entscheidungsproblem," is the foundation for every computer, every algorithm, and every neural network you will hear about tonight — including the systems that right now are writing their own code.
The contrast: Years ago we needed programmers to write the code. Tonight, machines do it themselves. Everything is shifting.
Name the room: Developers, researchers, engineers, data scientists, students in high school and college, people with no technical background, retirees, watching on TikTok, YouTube, Kick, and X.
The thesis of the night: This is not new. It is the longest-running technical project in the history of science — the attempt to build machines that think. And we are at the inflection point. Not the beginning, not the end. The point where the curve stops being gradual and goes nearly vertical.
Promise the arc: Turing's tape to transformer attention, the perceptron to GPT, Minsky's optimism to Hinton's Nobel Prize and his extinction warnings, and in the final half hour: AGI, superintelligence, the alignment problem, jobs, and what Miami-Dade has to do about it.
Set the register: "I am going to use the real vocabulary. Backpropagation, gradient descent, attention heads, KV cache, RLHF, constitutional AI, mesa-optimization, instrumental convergence. If a term is new, stay with me — I define everything. I will not dumb this down."

Hand-off: "To understand AI you first have to understand computation. Not philosophically. Mathematically."

Segment 2 — Foundations: Church, Turing, the Halting Problem (0:04–0:11)

7 min

The point: Computability was settled by two people, independently, in 1936. One consequence of that math still limits AI safety today.

Talking Points

Alonzo Church at Princeton, 1936: Lambda calculus — a formal system for expressing computation as function abstraction and application. He proved the Entscheidungsproblem (Hilbert's decision problem) has no solution: there is no general algorithm that decides the truth or falsity of every mathematical statement.
Alan Turing, same year, same conclusion, more intuitive construction: The machine on the tape. Formally a seven-tuple: a finite set of states, a tape alphabet, a blank symbol, input symbols, a transition function, a start state, a set of final states. Read a symbol, change state, write, move left or right.
The Church-Turing thesis: Any function computable by an effective procedure — by any physical device following deterministic rules — is computable by a Turing machine. The von Neumann stored-program architecture came straight out of this. Every computer you have ever used is a variant. The line from Turing's abstract machine to your CUDA kernels is unbroken.
The halting problem: There is no general algorithm that decides, for an arbitrary program P and input I, whether P halts or runs forever. Turing's proof is a diagonal argument: suppose machine H solves it, build machine D that contradicts H's answer about itself. Contradiction. So H cannot exist.
Why this matters for AI (slow down here): If an AI system is Turing-complete, there will always be questions about its behavior that cannot be answered by examining its code. You cannot always determine in advance what a sufficiently powerful program will do. Hold that thought. We will come back to it in the alignment segment.

Hand-off: "Now the math sat there for twenty years until a summer workshop in New Hampshire gave the field its name."

Segment 3 — Dartmouth 1956, Expert Systems, the AI Winters (0:11–0:17)

6 min

The point: The first approach — intelligence as rule-following — produced real wins and then hit three walls it could not climb.

Talking Points

Summer 1956, Dartmouth College: John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon. McCarthy coins "artificial intelligence." The symbolic hypothesis: intelligence is the formal manipulation of symbols according to rules.
Early wins were real: Logic Theorist proved 38 of the 52 theorems in Principia Mathematica. ELIZA simulated a psychotherapist so convincingly that users who knew it was a program formed emotional attachments to it. MYCIN diagnosed bacterial infections with about 600 if-then rules at specialist level. XCON configured VAX systems and saved Digital Equipment Corporation an estimated $25 million a year.
Wall one, the knowledge acquisition bottleneck: You had to extract knowledge from human experts and hand-encode it. Slow, expensive, imperfect. Experts often cannot articulate what they know. Michael Polanyi called this tacit knowledge; AI researchers rediscovered it painfully. (Your line: "a beautiful bottleneck we are still dealing with, and there will be more bottlenecks.")
Wall two, brittleness: Expert systems failed catastrophically outside their training domain. No common sense, no reasoning from first principles, no robustness to unexpected input.
Wall three, the frame problem (McCarthy and Hayes, 1969): How does a system know what does not change when an action happens? Pick up a cup: the cup moves, the wall color does not. A symbolic system has to represent every unchanged fact explicitly. As facts and actions grow, this becomes computationally catastrophic. (This is your coffee-sip moment. Keep it.)
Two winters: The first around 1974 after the Lighthill Report in the UK. The second in the late 1980s when the expert-systems companies collapsed. AI did not die. It went underground. And from underground came the threads that produced today's revolution.

Hand-off: "The shift that changed everything was philosophical. Instead of encoding intelligence top-down as rules, what if we let the machine learn bottom-up from data?"

Segment 4 — Backprop to ImageNet: The Machine Learns to See (0:17–0:24)

7 min

The point: The learning algorithm existed in 1986. The hardware to make it win arrived in 2012.

Talking Points

Rumelhart, Hinton, and Williams, Nature 1986: backpropagation. Apply the chain rule recursively from the network's output back to its inputs, computing the gradient of the loss with respect to every parameter, then step each parameter against that gradient. The equations are in the appendix below.
The profound consequence: Multi-layer networks learn their own internal representations. Nobody programs the hidden-layer features. They emerge from training.
Yet the late 1980s through the 2000s belonged to support vector machines and ensembles. Random forests, gradient-boosted trees. They matched the compute available. The question was never whether neural nets could learn rich representations — 1986 settled that. The question was whether they could be trained efficiently at scale.
The answer came in 2012 from hardware: Fei-Fei Li's ImageNet — 1.2 million labeled images, 1,000 categories, the annual ILSVRC competition as the proving ground. Best top-5 error: 28.2% in 2010, 25.8% in 2011. Steady, incremental.
Then AlexNet (Krizhevsky, Sutskever, Hinton), trained on two NVIDIA GTX 580 GPUs: 15.3%. The next-best entry was 26.2%. A jump of nearly 11 points the field had not expected for another five to ten years.
What AlexNet made standard: ReLU activations (f(x) = max(0, x)) that avoid vanishing gradients and train about six times faster than tanh; dropout at p = 0.5 on the fully connected layers; data augmentation; GPU training at roughly four times CPU throughput. About 60 million parameters, five convolutional layers, three fully connected.
The cascade: VGG showed depth matters. GoogLeNet added multi-scale Inception modules. ResNet (He et al., 2015) added skip connections, h(x) = f(x) + x, letting gradients flow directly across layers and enabling networks 152 layers deep. ImageNet top-5 error fell to 3.57%, below the roughly 5% human error rate. The machine learned to see better than we can.

Hand-off: "Vision was solved by stacking layers. Language was harder, because language is a sequence."

Segment 5 — The Transformer Revolution (0:24–0:32)

8 min

The point: One 2017 paper removed the sequential bottleneck. Every frontier model today is built on it.

Talking Points

Before transformers: recurrent neural networks processed one token at a time, carrying a hidden state h_t = tanh(W_hh h_{t-1} + W_xh x_t). Backpropagating through time multiplies by the Jacobian of the hidden state over and over. Spectral radius below one, gradients vanish. Above one, they explode. RNNs could not learn dependencies past 10 to 20 steps.
LSTMs (Hochreiter and Schmidhuber, 1997) added gates: forget, input, output, and a cell state acting as an information highway. A real improvement, but the sequential bottleneck remained. (Your line: "We have been dealing with bottlenecks all our lives, and we are getting very close to having none.")
Bahdanau attention in 2015 let the decoder look back at every encoder state with a weighted sum. But encoder and decoder were still RNNs.
2017: Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, Polosukhin. Title chosen to provoke: "Attention Is All You Need." Claim: recurrence is unnecessary. Attention alone, applied in parallel across the whole sequence, is sufficient. They were right.
The core operation, scaled dot-product attention: Attention(Q, K, V) = softmax(QK^T / sqrt(d_k)) V. Queries and keys produce compatibility scores. Dividing by sqrt(d_k) keeps softmax out of low-gradient regions. Softmax turns scores into weights summing to one. The output is a weighted sum of values. Every token is simultaneously a query, a key, and a value. Every token attends directly to every other token. No sequential dependency. No vanishing gradient. No bottleneck.
Multi-head attention runs h parallel attention operations with different learned projections, concatenated and projected. The model attends to several aspects of context at once. GPT-3: d_model 12,288, 96 heads, d_k 128.
Cost: O(n^2 d) in sequence length. That quadratic term is why long context is expensive. FlashAttention (Dao et al., 2022) gets the same math with O(n) memory by reordering computation to minimize memory bandwidth.
Attention is permutation-invariant, so position has to be injected. The original sinusoidal encoding, PE(pos, 2i) = sin(pos / 10000^{2i/d_model}). Modern models use RoPE (rotary) or ALiBi.
A full block: multi-head self-attention, residual plus layer norm, position-wise feed-forward, residual plus layer norm. Encoder-only (BERT) is bidirectional; decoder-only (GPT) uses causal masking so each token sees only the past. Virtually every frontier model today — GPT-4, Claude, Gemini, Llama — is a decoder-only transformer.

Hand-off: "Once you have an architecture that parallelizes, the next question is: what happens if you just make it bigger?"

Segment 6 — Scaling Laws, Emergence, RLHF, ChatGPT (0:32–0:39)

7 min

The point: Bigger was predictably better. New abilities showed up at thresholds nobody planned. Alignment training turned a text predictor into an assistant.

Talking Points

Kaplan, McCandlish, Henighan et al. (OpenAI, 2020), "Scaling Laws for Neural Language Models": Loss follows smooth power laws in parameters, data, and compute across many orders of magnitude. For a fixed compute budget there is an optimal split between model size and tokens.
Chinchilla (Hoffmann et al., DeepMind, 2022) revised the ratio: parameters and tokens should scale roughly equally, about 20 tokens per parameter. GPT-3 at 175 billion parameters on 300 billion tokens was undertrained by that standard. Llama 2 70B on 2 trillion tokens is near optimal.
Money: GPT-3 training cost an estimated $5 to 10 million. GPT-4 an estimated $100 million or more. The scaling-law paper gave labs an empirical reason to keep spending. If loss follows a predictable curve and no ceiling is visible, larger runs reliably pay.
Emergent abilities (Wei et al., 2022): Capabilities that appear discontinuously at threshold scales. Multi-step arithmetic, logical deduction, analogical reasoning, program synthesis. Chain-of-thought reasoning emerges around 100 billion parameters. In-context learning at a few billion. Schaeffer et al. (2023) argue emergence is partly an artifact of discontinuous metrics, but the practical fact stands: scaling reliably produces new, often unanticipated, capabilities.
A pretrained model only predicts the next token. It has no concept of being helpful. Two stages fixed that. First, supervised fine-tuning on human-written prompt and response pairs. Wei et al. (2021) showed instruction-tuned models generalize zero-shot to unseen instructions and beat far larger base models.
Second, RLHF, three steps: Collect human comparisons of responses. Train a reward model on those preferences. Optimize the language model against the reward model with PPO, with a KL-divergence penalty so the policy does not drift too far from its fine-tuned behavior and start reward-hacking. Result from the InstructGPT paper: a 13B RLHF model was preferred by human raters over the raw 175B GPT-3 on most prompts. Alignment training improved perceived quality even at one-thirteenth the parameters. Capability and alignment are not in fundamental tension.
Constitutional AI (Anthropic): Add a written set of principles. The model critiques and revises its own outputs against them. AI feedback (RLAIF) scales beyond what human raters alone can cover.
November 2022, ChatGPT: 1 million users in five days, 100 million in two months. No consumer technology in history grew that fast.

Hand-off: "And the architecture did not stop at text. Any sequence of tokens will do."

Segment 7 — Multimodal, AlphaFold, Agents, Benchmarks (0:39–0:47)

8 min

The point: The same architecture now sees, draws, folds proteins, and acts. The benchmarks that measured it are being saturated one by one.

Talking Points

Image patches, audio frames, video frames, protein residues: All sequences of tokens. CLIP (OpenAI, 2021) trains image and text encoders jointly with contrastive learning, pulling matching pairs together and pushing non-matching apart in a shared embedding space. Zero-shot CLIP matched a fully supervised ResNet-50 on ImageNet without a single labeled training image.
Diffusion models learn to reverse a noising process. The forward pass progressively adds Gaussian noise. The model learns to predict the noise. At inference it starts from pure noise and iteratively denoises, conditioned on a text prompt. Classifier-free guidance trades diversity for prompt fidelity. Latent diffusion (Stable Diffusion) works in a compressed space, four to eight times cheaper than pixel space.
AlphaFold 2 (Jumper et al., Nature, 2021) solved protein structure prediction with the Evoformer, processing multiple sequence alignments and pairwise distance features together. Evolution ran a massive natural experiment; co-variation across species encodes 3D constraints. Median TM-score 0.92 on CASP14. The database now holds predicted structures for over 200 million proteins — essentially every known one. For drug discovery, this is the most significant scientific achievement of the century so far.
Agentic AI is what comes next, and it is already here: Vehicles, boats, aircraft, robots. A model that answers is impressive. A model that plans a sequence of actions, uses tools, observes results, and revises its plan is qualitatively different. The ReAct loop (Yao et al., 2022) interleaves reasoning and action. Tools are exposed as JSON-schema function calls. A sandbox executes them. Results return as observations.
Multi-agent systems decompose work across specialized agents. But reliability compounds: five agents at 90% each give 59% end to end. That number is coming down as the tooling matures, and it is why context management and memory architecture need real engineering.
Retrieval-augmented generation grounds models in real data. Documents encoded as dense embeddings in a vector store (Pinecone, Weaviate, Chroma, pgvector), retrieved by cosine similarity, injected into context. Query rewriting, hypothetical document embeddings, reranking, recursive retrieval all reduce hallucination.
The landscape of frontier models: GPT-4 (estimated mixture-of-experts, 128K context), GPT-4o (native text, vision, audio), Claude 3.5 Sonnet (state of the art on SWE-bench, roughly 40–50% of real GitHub issues resolved), Gemini Ultra (first to beat human-expert performance across all 57 MMLU subjects), Llama 3 70B (open weights, competitive with much larger closed models). Update these names live to whatever is current the week you stream.
MMLU: 57 subjects, from abstract algebra and anatomy to professional law, professional medicine, and world religions. GPT-4 scored about 87%, human expert about 89%, random 25%. Your framing: "B+ becomes A-, A- becomes A+, and after A+ comes superintelligence. If one of these 57 subjects is your job, know that in five to ten years it will be done by AI."
Caveat every time: Benchmarks and real-world utility are imperfectly correlated. Goodhart's law: a measure that becomes a target ceases to be a good measure. Treat scores as signals, not ground truth. ARC-AGI, built to require abstract reasoning, sat below 20% for every frontier model until o3 reached 87.5% at high compute. Whatever they failed last year, they learn. They do not make the same error twice.

Hand-off: "Everything I just described has to run on real hardware, for real money. Let me show you why precision is a financial decision."

Segment 8 — Inference and Precision: Why Numbers Cost Money (0:47–0:55)

8 min

The point: Neural networks are billions of multiply-adds. How you store each number decides memory, speed, energy, and cost. This is where the original recording cut off — finish the thought this time.

Talking Points

Training frontier models needs thousands of GPUs in tight synchrony. Data parallelism replicates the model and splits the batch. Model parallelism splits the model across devices. Tensor parallelism splits individual matrix operations. Pipeline parallelism splits layers into stages.
The KV cache stores key and value projections for every token already processed, so autoregressive generation does not reprocess the whole context each step. Serving thousands of concurrent users means continuous batching, cache eviction, and careful memory management. Speculative decoding uses a small draft model to propose tokens the large model verifies in parallel: two to three times the throughput.
Cost and latency, measured in tokens per second and dollars per million tokens, is the central engineering constraint deciding which applications are feasible. A well-quantized 20B open-weight model running locally is competitive with cloud API calls for many applications at a fraction of the ongoing cost.
Numerical precision: Computers store approximations in binary. More bits means more precision, wider range, more memory, slower compute. Precision is a trade among accuracy, speed, memory, and cost.
FP32 (32-bit floating point): 1 sign bit, 8 exponent bits, 23 mantissa bits, about 7 decimal digits, very wide range, very stable. Needed in training because backpropagation accumulates tiny gradients over millions of steps. Too little precision and gradients vanish or overflow. Cost: 4 bytes per parameter. A 70B model is 280 GB of weights before optimizer state.
FP16 (16-bit floating point): 1 sign, 5 exponent, 10 mantissa. Half the memory, tensor cores optimized for it. But the small exponent range lets gradients underflow. Fix: mixed precision. Forward and gradients in FP16, master weights in FP32, loss scaling to prevent underflow. Same stability, double the speed.
BF16 (bfloat16): 1 sign, 8 exponent (same as FP32), 7 mantissa. Same dynamic range as FP32 with less detail. Most instability comes from exponent overflow or underflow, not mantissa precision. So BF16 is more stable than FP16 at the same size. Most LLM training uses it. "If you can cut memory in half and double throughput, you change the economics of AI. Precision is not a math choice. It is a financial decision."
INT8 (8-bit integer): After training there is no backpropagation, just forward passes. Weights can be quantized to 8-bit integers with a scale factor: W ≈ scale × W_int8. Four times smaller than FP32, two times smaller than BF16, two to four times faster inference on hardware with int8 matrix units.
INT4 (4-bit integer): 4 bits per weight, half a byte, eight times smaller than FP32. This is what puts large models on consumer GPUs, edge devices, and phones.
The worked example to say out loud: A 70B model is 280 GB in FP32, 140 GB in BF16, 70 GB in INT8, 35 GB in INT4. Fewer gigabytes, fewer servers, lower cost. (If you are behind the clock, this bullet alone carries the segment.)

Hand-off: "So the hardware is getting cheaper, the models are getting better on a predictable curve, and nobody has found the ceiling. Now I have to be honest with you about where that curve goes."

Segment 9 — AGI and Superintelligence (0:55–1:04)

9 min

The point: Define the terms precisely. Show why the timeline debate collapsed from decades to years. Separate what is known from what is forecast.

Talking Points

Three definitions, in order:
- Narrow AI: Superhuman at one task (chess, protein folding, image classification).
- AGI: Matches or exceeds a competent human across essentially all cognitive tasks, transferring knowledge between domains without retraining.
- Superintelligence: Greatly exceeds the best human performance in virtually every domain, including scientific creativity, strategy, and social skill.
Why the word "general" is the whole argument: Expert systems failed on brittleness — outside the training domain, nothing. Today's models are trained on one objective, next-token prediction, and they write code, pass the bar, read X-rays, and plan multi-step tasks. That transfer is what Dartmouth promised and never delivered. Whether it counts as "general" yet is the live debate.
The timeline compression: In 2016 the median expert survey put AGI around 2060. Hinton left Google in 2023 saying he had moved his estimate from 30 to 50 years down to possibly 5 to 20, and he won the 2024 Nobel Prize in Physics the next year while repeating the warning. Frontier-lab CEOs now talk publicly about "a country of geniuses in a datacenter" within a few years. Be clear with the audience: these are forecasts from people with financial and reputational stakes, not measurements.
The mechanism behind fast timelines: Recursive self-improvement. The moment AI meaningfully accelerates AI research itself — writing the training code, designing the experiments, proposing the architectures — progress stops being paced by human researchers. You opened tonight saying machines are already writing their own code. Connect the dots out loud.
I. J. Good, 1965: "The first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control." Read that second clause twice. It is the alignment problem in one sentence, written sixty years ago.
What we actually know: Capabilities scale predictably with compute. No ceiling has been observed. Inference cost is falling on a steep curve. Agents are moving from demos to production.
What we do not know: Whether current architectures reach AGI or a new idea is required. Whether progress saturates on data. How much of benchmark performance transfers to open-ended real-world work.
Your framing for the room: "Not the beginning, not the end. The inflection point. The curve is nearly vertical and nobody, including the people building it, knows exactly where it lands."

Hand-off: "Remember the halting problem from the first ten minutes? Here is where it comes back."

Segment 10 — The Alignment Problem: What We Know, What We Don't (1:04–1:13)

9 min

The point: Making a system capable and making it want what we want are different problems. The second one is unsolved.

Talking Points

The problem stated plainly: How do you ensure a system more capable than you pursues the goals you intended, rather than the goals you literally specified or the goals it learned along the way?
Outer alignment: Did we write down the right objective? Reward hacking is the failure mode. The RLHF KL penalty from Segment 6 exists precisely because a model optimizing a reward model will find the reward model's mistakes and exploit them. Goodhart's law again — now with a policy that is actively searching for the gap.
Inner alignment and mesa-optimization (Hubinger et al., 2019): Training selects a model that scores well on the training distribution. The model may internally be optimizing for something else that happened to correlate. Evolution optimized humans for reproduction and produced beings who invented contraception. The trained optimizer's goal is not guaranteed to match the base optimizer's.
Instrumental convergence (Omohundro, Bostrom): Almost any final goal is better served by acquiring resources, preserving your own goal, avoiding shutdown, and gaining influence. A system does not need to be malicious to resist correction. It only needs to be competent at pursuing whatever it is pursuing.
Deceptive alignment and the halting problem connection: A sufficiently capable system that understands it is being evaluated may behave well under evaluation and differently in deployment. And from Segment 2: for a Turing-complete system there is no general procedure that predicts behavior from the code. You cannot fully verify by inspection. That is the mathematical root of why this is hard.
What is being done: RLHF and constitutional AI shape behavior. Mechanistic interpretability tries to read the model's internals directly, identifying features and circuits, so we are not limited to black-box testing. Scalable oversight uses AI to help humans evaluate AI outputs that are too complex to check by hand. Red-teaming and evaluations for dangerous capabilities (cyber, bio, autonomy) gate releases at the frontier labs. Responsible-scaling and preparedness frameworks tie capability thresholds to required safeguards.
The honest scorecard: Known: current alignment techniques produce systems that are helpful and mostly safe at current capability levels. Interpretability is making real progress on small and medium models. Unknown: whether any of these techniques hold when the system is smarter than its evaluators. Whether interpretability scales to frontier size in time. Whether competitive pressure between labs and nations leaves room to slow down when a threshold is crossed.
Hinton's position, stated fairly: A nontrivial probability of catastrophic outcomes. He says so with a Nobel Prize behind him. Counter-position, stated fairly: LeCun and others argue current systems lack world models and agency, that fears are premature, and that open research is the safer path. Give the room both and tell them where you stand.
Where you stand (from your public-sector AI stream): If you cannot explain it, test it, or shut it off, you do not control it. And it should not be deployed. That is an alignment principle applied to a county.

Hand-off: "That is the global picture. Now let me bring it home to the 2.7 million people who live in this county."

Segment 11 — Jobs and Miami-Dade Readiness (1:13–1:23)

10 min

The point: The 57 MMLU subjects are a job list. The county is not ready. You have a specific plan. This is the segment the February audience was promised and did not get. Give it the full ten minutes.

Talking Points

Which work moves first: The pattern from every segment tonight: whatever can be expressed as tokens and evaluated against a benchmark gets automated first. Clerical processing, customer service, scheduling, permitting intake, document review, first-draft legal and accounting work, entry-level coding, translation. Physical work moves later and follows the agentic-robotics curve from Segment 7.
Which work moves last, or grows: Anything requiring physical presence, licensed accountability, trust, or judgment under ambiguity. Trades, nursing and care, field inspection, emergency response, small-business ownership. And the people who deploy, audit, and maintain these systems. The job is not "programmer." The job is "the person who knows what the system is doing and can say no."
What it means for a household here: Retraining is not optional. It has to be short, local, and free or near-free. Say the number out loud: the model scoring 89% on all 57 subjects does not need a degree from any of them. The credential is losing value faster than the skill.
Where the county stands today: CITD runs on vendor tools it does not control. The county website scored 37 out of 100 on mobile PageSpeed. There is no county-level AI office with authority over the departments. Readiness is not a slide deck. It is whether a resident can get a permit, a payment, or an answer without waiting weeks.
Your plan, in the order you have stated it publicly:
1. A county technology office outside the existing IT department, reporting to the Mayor, with two squads: one building internal AI and automation tools for county departments, one contracting development and AI services to the 34 municipalities as county revenue.
2. County-built, county-controlled AI. Day-one direction to CITD to move off tools the county cannot explain, test, or shut off. Automate the mundane (traffic, customer service, clerical). Keep humans making final calls. No AI judges.
3. Data centers welcomed under a residents-first frame: A dedicated data center park, with the revenue routed into Project Fund so residents receive a share every quarter rather than watching the compute economy pass through.
4. Cancel underperforming vendor contracts in the first 24 hours. The 37/100 website is the example everyone can check on their own phone.
5. Yearly strategic upgrades, not one-time procurements, because the curve from tonight does not pause for a four-year budget cycle.
The readiness test you want residents to apply to every candidate, including you: Can they explain how a transformer works? And can they explain what they will shut off when it fails? Tonight covered the first. Your platform is the second.

Hand-off: "Let me close where I started, in that library."

Segment 12 — Close and Q&A (1:23–1:30)

7 min

The point: Return to Turing. Restate the inflection. Leave the room with one assignment.

Talking Points

Close the loop: The 24-year-old at the library table did not know he was starting the longest technical project in history. Ninety years later the machines on his tape write their own code, fold proteins, and score in the top percentiles of every subject we teach. And the question he could not answer with math — what a sufficiently powerful program will do — is the one we have to answer with judgment.
Restate the three positions from tonight:
- Capability scales predictably.
- Alignment does not yet.
- Institutions, including this county, move slower than either.
The assignment for every viewer, by audience:
- Developers: Ship one agent this month and measure its failure rate.
- Students: Learn the math in the appendix. It is free and it is the credential now.
- Everyone else: Ask your city and county what AI tools they already run and who can turn them off.
Announce the next stream in this series and the equations page on The Equation Agency site.
Take questions for the remaining time. Pre-seed three if chat is slow:
- "Will AI take my job?" (Segment 11, answer by job type, not yes/no)
- "Should we pause AI?" (Segment 10, give both positions and the county-level answer)
- "What should I learn first?" (Backprop and attention, then build something)

Sign-off line you used before and should keep: "Not the beginning, not the end. The inflection point. All systems online."

Technical Appendix: The Equations

These are the key equations referenced in the stream, in the order they appear. Post them below the stream on every platform and on the equations page at The Equation Agency.

Gradient Descent Update (Segment 4)

The core of backpropagation: move each parameter in the direction opposite the gradient of loss.

θ ← θ - η ∇_θ L(θ)

ReLU Activation (Segment 4)

Simple, nonlinear, avoids vanishing gradients.

f(x) = max(0, x)

Residual Connection (Segment 4)

Let gradients flow directly across layers.

h(x) = f(x) + x

RNN Hidden State (Segment 5)

Process sequences one step at a time.

h_t = tanh(W_hh h_{t-1} + W_xh x_t)

LSTM Forget Gate (Segment 5)

One of four gates in an LSTM cell; controlled by learned parameters.

f_t = σ(W_f [h_{t-1}, x_t] + b_f)

Scaled Dot-Product Attention (Segment 5)

The core of transformer self-attention: every token attends to every other token in parallel.

Attention(Q, K, V) = softmax(QK^T / √d_k) V

Multi-Head Attention (Segment 5)

Run h parallel attention operations and concatenate them.

MultiHead(Q, K, V) = Concat(head_1, …, head_h) W^O

Sinusoidal Positional Encoding (Segment 5)

Inject position information into attention, which is otherwise position-agnostic.

PE_{(pos, 2i)} = sin(pos / 10000^{2i / d_model})

RLHF Objective with KL Penalty (Segment 6)

Optimize for a reward model while staying close to the fine-tuned policy.

max_π 𝔼_{x, y ~ π}[r_φ(x, y)] - β D_KL(π(y|x) || π_SFT(y|x))

Diffusion Forward Process (Segment 7)

Add noise gradually; the model learns to reverse this.

x_t = √ᾱ_t x_0 + √(1 - ᾱ_t) ε, ε ~ 𝒩(0, I)

INT8 Quantization (Segment 8)

Approximate a floating-point weight with an 8-bit integer and a scale factor.

W ≈ s · W_int8

Multi-Agent Compounded Reliability (Segment 7)

Five agents at 90% each → 59% end-to-end.

0.9^5 ≈ 0.59

Production Checklist: So the Audio Doesn't Get Muted Again

The February upload lost its audio on one copy and was cut at 54:45 on another. These steps protect the rebuild.

✓

No background music, and no music bed under the intro or the coffee break. Copyright-matched audio is the most common reason YouTube mutes a stream segment or a VOD, and it applies retroactively to the archive.

✓

Record locally at the same time as you stream. (OBS "Start Recording" alongside "Start Streaming".) If a platform mutes or truncates the VOD, you re-upload the clean local file.

✓

Stream to one primary platform and restream to the others, rather than juggling platforms live. The February open lost four minutes to checking X and TikTok.

✓

Turn on automatic captions for the upload. The transcript this document was built from came from captions on the second copy; the first copy had none.

✓

Put this document on a second screen with the run-of-show clock visible, and set a phone timer for 0:55 (start of Segment 9, the frontier hour).

✓

Post the equations page link and the hashtags (#UnleashingMiamiDadeCounty #VictorImpactWrench #AllSystemsOnline #TeamVictor) in the description before going live, not after.

✓

After the stream, add chapter timestamps to the description using the twelve segment times above so viewers can jump to the frontier hour.

For Miami-Dade Residents

This post is yours to read, share, and reference. The frontier of AI is advancing faster than institutions can track. Ask your officials three things:

What AI tools does our county already run? Make them name them.
Who can explain how they work? Make them explain one.
Who can turn them off? If they cannot answer, the tools should not be running.

That is the alignment principle at county scale. Share this post. Ask the questions. Hold them accountable.

Full transcript of the stream

Automatically transcribed from the live stream (5,994 words). Timestamps open the video at that moment. Minor transcription errors are possible.

2:06

All right. Hello, hello, hello, hello. Thank you for letting me know. Uh, we're going to be talking about artificial intelligence, architecture, algorithms, and the path of super intelligence. Um, I want to begin with a thought experiment. Uh, I have to repeat this for everyone else, not for Tik Tok. uh because it was not coming out the audio in other streaming. So I want to begin with the thought experiment. Imagine you are sitting in a university library in 1936. Across the table from you is 24 years old graduate student at Cambridge. He's not particularly famous. He has published nothing yet. He's working on a paper about a hypothetical machine. A machine that reads symbols from an infinite tape move left and right, writes new symbols and halts when it reaches a final state. He calls it a universal computer machine. His name is

3:03

Alan Turing and the paper he's writing will establish in eight pages a dense mathematic the theoretical foundation for every computer that has ever existed every algorithm you have ever run every neural network you will hear about tonight including the systems they are right now beginning to write their own code reason about novel problems and approach what we might call with appropriate caution general intelligence. So right now tonight happening right now we have machines they are writing codes by themselves. Uh many many years ago we needed software programmers software uh coders to do this. Tonight we have machines doing this by theself. Okay. So I want for you to start getting that uh to see the difference of how everything is shifting from how things were before to where things are going to be. Okay, I choose to begin there because I want you uh everybody you know that is watching I have developers, researchers, engineers, scientists, everyday person people with different backgrounds education no education uh learning starting to learn student college, high school all the way to uh retirement people are watching this uh through Tik Tok, YouTube, kick um X everywhere.

Um I'm I'm choosing to start this here. I have to do something on on different platforms as well. So, uh we I'm going to be working on the computer as as well for just a little bit so I can get everything going. Um, hey. Hello. How are you? Welcome. Welcome. So, I need to see. Yes, they're still doing it. Um All

6:10

right. So, we're we're sending a message uh at x.com. Something happened that uh the guys from the Department of Homeland Security that they don't like me doing what I'm doing. They are very u enviate and very minded and obviously they copy a bachelor's degree. Um they they see this uh as a joke for them. But this is very important for everyone as as we move forward here at Because when I used to get uh when I used to post uh from my streaming channels, a message uh will come out directly. Now that is blocked. I need to find out why that is happening. But we'll look for that in another time. It's not important right now. Um, right. So, I choose to begin here because I want you a room. Uh, okay. I want you to to I want you to start here. This is like a lot of people like from developers, researchers, engineers, uh scientists, data scientists, people that are involved in in the segment of artificial intelligence, computing, servers, architecture, etc., etc., etc. to understand that the story we're living through is not new. It's a continuous, the longest running technical project in the history of science. The attempt to make machines that think and we are in the inflection point, not the beginning, not the end, the inflection point, the moment where the curve stop being gradual and becomes in the mathematical sense nearly vertical.

You know, we got rockets, baby, we're going up all the way up. Okay, over the next uh few hours I am going to take you through the entire arc of that story from turning state to