The AI Revolution Is Underway · Training Deck
AI Literacy & Hands-On Tools Training · July 2026

The AI Revolution Is Underway

See the big picture  ·  Master the tools  ·  Build it yourself

← → or Space to navigate  ·  F for fullscreen  ·  47 slides

Today's Roadmap

90 minutes to see an entire revolution clearly

01 A Real Revolution

How fast is AI actually improving? And what did Elon Musk really predict?

02 A Bet of National Destiny

How the US and China are staking their futures on AI as national strategy

03 The Battle of Models

US and Chinese models in close combat: how much of a gap is left?

04 Capital Markets

AI companies newly listed and lining up to list · China's "Six Little Tigers" diverge

05 AI Agents

How AI does the work for you: Claude Code, the "Lobster," and Hermes

06 Embodied AI

When intelligence gets a body: AGI, humanoid robots, dexterous hands

07 Hands-On Practice

A set of exercises to put today's lessons to work immediately

08 The Pro Workflow

Prep kit → iterate → review → ship & operate

01

A technological revolution that is really happening

Not hype. Not a story spun for investors. To judge whether a revolution is real, look at three things: how fast it spreads, how fast it improves, and where the real money goes.

01 · A Real Revolution · Putting AI on history's map

The fourth time around: every revolution has reshuffled the ranks of great powers

Ⅰ

Steam Revolution 1760s–

Steam engines · Textile machines · Railways

For the first time, machines replaced human muscle

Front-runner: Britain — the "empire on which the sun never set"; left behind: a Qing dynasty closed to the world

Ⅱ

Electric Revolution 1870s–

Electricity · Internal combustion · Assembly lines

Energy and manufacturing at scale — modern industry takes shape

Front-runners: the US and Germany — new powers rise, Britain's hegemony loosens

Ⅲ

Information Revolution 1950s–

Computers · The internet · Smartphones

Machines take over computing and moving information

Front-runner: the US locks in tech supremacy; China catches the last train and pulls even at the application layer

Ⅳ

Intelligence Revolution 2020s · happening now

Large models · Agents · Robots

For the first time, machines replace the human brain — cognition, judgment, creativity

Front-runners: the US and China, neck and neck — exactly why both have bet their national destiny on it

Two things are different this time: the first three revolutions replaced muscle and hands — this one replaces the brain itself, so there is no ceiling on its reach; and the first three took decades to spread, while this one spreads in months — see the next slide.
01 · A Real Revolution

No technology in history has ever spread this fast

Time to reach 50 million users (years)

2 months
Time for ChatGPT to reach 100 million monthly active users (Nov 2022 – Jan 2023) — the fastest-growing consumer app in history at the time
TikTok took 9 months; Instagram took two and a half years

Sources: Statista / HBR, "The Pace of Technology Adoption is Speeding Up"; UBS analysis (Reuters, Feb 2023)

01 · A Real Revolution

The point isn't that AI is powerful — it's that it's getting powerful faster and faster

4–5×
How much frontier-model training compute grows every year — a trend sustained for over a decade
Epoch AI, 2010–2024 trend
7 mo → 4 mo
How long it takes for the length of tasks AI can complete on its own to double — the doubling time itself is shrinking
METR time-horizon study, updated Jan 2026; current frontier models can handle tasks of roughly 12–14 hours
1.7% → 80%+
SWE-bench (real software-engineering tasks) scores: from GPT-4 (2023) to 2026 frontier models — the benchmark is nearly saturated and being retired
Epoch AI / OpenAI, 2026
A landmark moment: in July 2025, general-purpose models from OpenAI and Google DeepMind both reached the International Mathematical Olympiad (IMO) gold-medal bar — something only 11% of the 630 human contestants managed that year.

Sources: epoch.ai/trends · metr.org/time-horizons · Axios 2025-07-21 · official DeepMind blog

01 · A Real Revolution · In the words of the people steering it
"AI that is smarter than any human by the end of this year… by 2030 or 2031 — about five years from now — AI will be smarter than all of humanity collectively." — Elon Musk, World Economic Forum, Davos, 2026-01-22
That same month he wrote on X: "We have entered the Singularity." And again on June 21: AI will probably exceed the combined intelligence of all humans "in 4 to 5 years."

Sam Altman · OpenAI CEO 2025.06

"We are past the event horizon; the takeoff has started." — from his blog post The Gentle Singularity

Dario Amodei · Anthropic CEO 2024.10–2025

"Powerful AI" could arrive as early as 2026–2027 — the equivalent of "a country of geniuses in a datacenter." — Machines of Loving Grace and Davos interviews

Sources: Rev Davos transcript · Euronews 2026-01-22 · X@elonmusk 2026-06-21 · OpenAI blog "The Gentle Singularity" 2025-06 · Anthropic "Machines of Loving Grace" 2024-10

01 · A Real Revolution

Real money doesn't lie

$725B
Combined 2026 AI capital expenditure of Microsoft, Google, Meta, and Amazon — a 77% surge over 2025
Analysts expect it to top $1 trillion in 2027
500×
Anthropic's annualized revenue growth in two and a half years: $87M in early 2024 → over $44B by May 2026
Official Series H figures; now ahead of OpenAI (~$25B)
48%
Share of Q1 2026's 78.6K tech layoffs attributed to AI/automation — the structure of employment is already being rewritten
Entry-level hiring is hit hardest; a counterexample: IBM tripled its junior hiring
The takeaway: fastest adoption ever, acceleration still increasing, and money entering by the trillions — this is not another hype cycle. It's the early stretch of an exponential curve. You are standing where the curve still looks flat.

Sources: Tom's Hardware / CNBC Feb 2026 · VentureBeat Apr 2026 · Goldman Sachs Research · CBS News

01 · A Real Revolution · What this revolution means for individuals

AI gives ordinary people angel wings

New ceiling: ? Old ceiling: set by years of training An ordinary person × AI

The old ceiling was set by years of training. The new one is set by imagination.

For the first time, one person can "grow" a whole team

Copywriting, coding, design, data analysis, law, foreign languages — each used to take years of training. Now all it takes is being able to say clearly what you want and what counts as good. The barrel effect has been flipped: your shortest stave used to set your floor — now AI fills in every weakness for you, and all you need is one plank that's long enough.

"The computer is a bicycle for the mind."
— Steve Jobs, 1980. Forty-plus years later, AI has traded the bicycle for wings.

"Everyone is now a programmer — the programming language of the future is human language."
— Jensen Huang, NVIDIA CEO, Computex 2023

"There won't be a profession called 'programmer' anymore — anyone who can speak will have a programmer's powers."
— Robin Li, Baidu CEO, 2024

02

A Bet of National Destiny: the US and China both go all in

Rarely in history has a single technology made the world's two largest economies place their bets at the same moment, with the full force of national will. This time, both sides have wagered their futures.

02 · A Bet of National Destiny · United States

Compute, energy, export controls: total war

America's AI Action Plan 2025.07

Winning the Race: 90+ federal actions on three pillars — accelerate innovation (cut regulation), build AI infrastructure (fast-track permits for data centers and the grid), and lead the international AI order (export a "full-stack AI package" to allies).

Stargate Target: $500B

OpenAI × SoftBank × Oracle × MGX. By the end of 2025, over $450B committed and more than 8 GW of compute planned; the ~1.2 GW flagship campus in Abilene, Texas is already online.

Energy as national power 2025–2026

Executive orders fast-track federal permits for data centers; the Department of Energy's "Genesis Mission" plugs AI into 17 national labs, aiming to "double American research productivity within a decade"; by 2030, data centers will account for 25% of new US electricity demand.

The chip-control tug-of-war with China Weaponized

Apr 2025: H20 sales banned → Aug: an unheard-of deal — hand over 15% of sales revenue for a license → China pivots to domestic chips and the H20 goes cold → Dec: H200 approved case by case. Chips have become the bargaining chips of the two powers.

Sources: whitehouse.gov 2025-07-23 · OpenAI official announcement 2025-09 · energy.gov (Genesis Mission, 2025-11-24) · NPR/Fortune 2025-08 · BIS 2025-12

02 · A Bet of National Destiny · China

The "AI+" Initiative goes on the national agenda

The State Council's "AI+" Initiative 2025.08

Three steps: by 2027, smart devices and agents reach 70%+ adoption → by 2030, over 90%, with the intelligent economy as a major growth engine → by 2035, China fully enters the intelligent society.

15th Five-Year Plan + the "intelligent economy" 2025.10–2026.03

AI is positioned as the main battlefield of "new quality productive forces"; the 2026 Government Work Report introduced the term "intelligent economy," deploying ultra-large compute clusters and compute–electricity coordination. China's core AI industry already exceeds RMB 1.2 trillion.

Compute self-reliance Ascend roadmap

Huawei published its first three-year chip roadmap: 950PR (Q1 2026) → 970 (2028), with in-house HBM; per Reuters, new state-funded data centers may only use domestic AI chips. State capital has poured over $100B into AI data centers.

Open source as national strategy The DeepSeek effect

Openness is written into the "AI+" policy itself. As of August 2025, Chinese open-weight models took 17.1% of global downloads, overtaking the US (15.86%) for the first time.

Sources: gov.cn 2025-08-26 · Xinhua/People's Daily Online 2026-03-13 · Huawei Connect 2025-09 · Reuters 2025-11 · USCC report 2026-03

02 · A Bet of National Destiny

Two playbooks, one astonishing result

23×
US private AI investment in 2025, as a multiple of China's
($285.9B vs $12.4B)
BUT
2.7%
Benchmark performance gap between the strongest US and Chinese models
(in May 2023 it was still 17.5–31.6 percentage points)
United StatesChina
Core playbookCompute + capital + closed-source frontierEngineering efficiency + open-source ecosystem + real-world deployment
Role of the stateDeregulation, fast-track permits, export controlsTop-level planning, government guidance funds, adoption targets
Notable models, 20255030 (doubled from 15 the year before)
Open-source download share15.86%17.1% (first-ever lead)

Sources: Stanford HAI, 2026 AI Index Report · USCC Mar 2026. Note: China's government guidance funds and other implicit investment are not counted under "private investment."

03

Close combat: the US-China battle of large models

Before January 2025, the consensus was "China is two years behind." Then DeepSeek made its move.

03 · The Battle of Models · Fundamentals I: What is AI

What a large model really is: a giant function that learned to finish sentences

Input "The capital of China is _" A trillion-parameter neural network (Transformer) Output: next-word probabilities Beijing 92% Shanghai 3% Tokyo 1%

It doesn't look up answers — it computes probabilities. Every bit of "intelligence" you see comes from predicting the next word, over and over again.

Parameters = a trillion knobs

A model is a vast mathematical network, and its "parameters" are the adjustable knobs inside — frontier models have over a trillion. Training is simply turning every knob to the right position.

Tokens = the building blocks of text

Models chop text into small chunks for processing (the Chinese word for "artificial intelligence" ≈ 2 tokens). APIs bill by the token — that's where the "per million tokens" on the pricing slide comes from.

Emergence = quantity becomes quality

Scaling laws: grow parameters, data, and compute together, and abilities like understanding, reasoning, and creativity "grow out on their own" — which is exactly why the whole world is scrambling for compute.

In one sentence: AI isn't "programmed" into being — it's "fed" into being. From oceans of text, it learned the patterns of language, and of the world itself.
03 · The Battle of Models · Fundamentals II: How training works

"Forging" a large model: first read the entire internet, then teach it to behave

① Pre-training

Read trillions of tokens of text (≈ a few million libraries), learning the patterns of language and the world through a guess-the-next-word game. This is where the compute and money get burned.

② Post-training

Human demonstrations + ratings (RLHF and friends) teach it to be helpful, honest, and well-behaved — from "can chat" to "can work."

③ Inference (usage)

Every question you ask is one "inference," billed by the token. Train once, serve hundreds of millions.

ModelGPU scaleDurationPublished cost figureWhy it matters
GPT-4 · 2023~25,000 A100s (SemiAnalysis estimate)~3 monthsAltman himself: "more than $100 million"Kicked off the large-model arms race
Llama 3.1-405B · 202416,000 H100s · 30.84M GPU-hours (official)~3 monthsNo figure disclosed (electricity alone runs tens of millions of dollars)The strongest open-source model of its day
DeepSeek-V3 · 20242,048 H800s · 2.79M GPU-hours~2 months$5.58M (official paper; GPU rental for the final run only)An efficiency revolution — it set off the "DeepSeek moment"
2026 frontier (GPT / Claude / Grok)Clusters like xAI's Colossus have reached 200,000 GPUsMonthsAmodei: single training runs are heading toward the $10 billion markEscalated into a nation-scale compute race
Read this bill and the first two chapters snap into focus: training cost ≈ chips × electricity × time — which is why the US and China fight over chips and power. It also explains why DeepSeek shook the world: it trained a 90-point model for pocket change.

Sources: Wired 2023 (Altman interview) · SemiAnalysis 2023-07 · Meta Llama 3.1 model card 2024-07 · DeepSeek-V3 technical report 2024-12 · x.ai · Amodei podcast interview 2024-06. Methodologies differ; figures are order-of-magnitude reference only.

03 · The Battle of Models · Team USA (flagship models on sale, July 2026)

The closed-source frontier: five giants slug it out

OpenAI · GPT-5.5 2026.04

The default model across all of ChatGPT; the GPT-5.6 line (Sol/Terra/Luna) is already in preview. Annualized revenue around $25B.

Anthropic · Claude Fable 5 2026.06

A brand-new Mythos-class model beyond Opus. Three days after launch, the US government suspended access for foreign users for 19 days on national-security grounds — the model itself became a geopolitical event. The budget agent tier, Sonnet 5, followed shortly after.

Google · Gemini 3.1 Pro 2026.02

First place in 12 of 18 benchmarks at launch; 3.5 Flash leads on speed, and 3.5 Pro is on the way.

xAI · Grok 4.3 2026.04

Native video input and 1M context; the rumored 10-trillion-parameter Grok 5 is still in training. xAI has been folded into the SpaceX empire.

Meta · Muse Spark 2026.04

Meta's first closed-source, API-only frontier model — the one-time open-source standard-bearer changes sides, and the "Llama era" is effectively over.

What they share: all closed-source, stacking compute like mad, charging by the token, chasing the frontier.

Sources: openai.com · CNBC 2026-06-09/30 · Forbes 2026-06-16 · deepmind.google · x.ai · VentureBeat 2026-04

03 · The Battle of Models · Team China (main models, July 2026)

The open-source legion: cheap, fast, good enough — and getting stronger

DeepSeek V4 Open source

1.6T parameters (only 49B active); inference compute cut to 27% of the previous generation. Foreign press verdict: "nearly frontier, at a fraction of the price."

Alibaba Qwen3.7-Max Open + closed

Trillion-parameter MoE; the Qwen family has spawned over 113,000 derivative models — more than Google and Meta combined.

ByteDance Doubao Seed 2.0

Built for hundreds of millions of users, with inference cost cut 90%; ranked No.1 in China on SuperCLUE's Chinese benchmark, Mar 2026.

Zhipu GLM-5.2 MIT open source

754B parameters, pointedly released the very night Fable 5 was suspended, under the slogan: "Frontier intelligence belongs to everyone."

Moonshot Kimi K2.6 Open source

A big leap in coding ability; the 2.5T+ parameter K3 is slated for Q3.

MiniMax M3

59% on SWE-bench Pro, claiming to beat GPT-5.5; per-token compute is 1/20 of the previous generation.

Baidu ERNIE 5.1

2.4T natively multimodal; officially claims leading performance at 6% of the pre-training cost of its peers.

Tencent Hunyuan

Fast-slow fused MoE, 295B total with 21B active, aimed squarely at agent workloads.

Sources: DeepSeek API announcement 2026-04-24 · simonwillison.net · Alibaba Cloud Summit 2026-05 · Caixin 2026-02 · OSCHINA 2026-06 · official ERNIE blog 2026-05

03 · The Battle of Models

How much gap is left? Measure with three rulers at once

LMArena global blind-test leaderboard (Elo, 2026-07-01) — the top 10 are all American models, but the pack behind is closing in

USChina

The leader and China's best are only 34 Elo points apart; by Stanford's AI Index measure, the performance gap is 2.7%; by Epoch AI's time-based measure, China trails by about 7 months. Note: GPT-5.6 and GLM-5.2 both shipped in June — GPT-5.6 is still in closed preview and not yet on the blind-test board; adjacent models' score gaps overlap within confidence intervals.

Rewind to the "DeepSeek moment" (2025-01-27): DeepSeek disclosed that V3's final training run cost only about $5.6M in GPU time (mind the scope: excludes R&D and hardware). That day NVIDIA shed $589 billion in market value — the largest single-day loss in US stock-market history. The "China is two years behind" narrative died on the spot.

Sources: arena.ai/leaderboard (captured 2026-07-01) · Stanford HAI 2026 · epoch.ai · CNBC/Bloomberg 2025-01-27

03 · The Battle of Models

The pattern is set: America rules the frontier, China owns the ecosystem

API input price (USD per million tokens, official list prices, July 2026)

USChina
Claude Fable 5
$10.00
GPT-5.5
$5.00
Claude Sonnet 5
$2.00
Zhipu GLM-5.2
$1.40
DeepSeek V4-Pro
$0.44
DeepSeek V4-Flash
$0.14

Chinese open-source flagships list anywhere from a fraction of US flagship prices (GLM-5.2) all the way down to 1/35 – 1/100 (DeepSeek) — those short orange bars are why developers worldwide are voting with their feet.

41%
Share of global Hugging Face model downloads over the past year that went to Chinese models; the most popular open model is now DeepSeek-R1, dethroning Llama
Official Hugging Face report, Mar 2026
Top 10: all closed
LMArena's top ten is swept by closed American models; after Meta abandoned Llama for closed source, the open-source field was essentially ceded to China
arena.ai · VentureBeat Apr 2026

Sources: DeepSeek / Anthropic / z.ai official pricing pages (captured Jul 2026) · GPT-5.5 pricing via third-party trackers · huggingface.co spring report 2026-03-17 · SCMP

04

Capital markets: where the money flows

In 2026, AI companies are stampeding toward the exchanges — the US is incubating the largest IPO in history, while China's model makers rushed to ring the bell first.

04 · Capital Markets · US · Listed

The throne and the newcomers

NVIDIA NVDA The throne

≈ $4.8T

The most valuable company on Earth. The shovel-seller is still the biggest winner.

Cerebras CBRS IPO 2026.05

$5.5B raised

The AI-chip challenger and the biggest tech IPO of 2026 so far; priced at $185, up as much as 108% on day one. Holds a $10B compute contract with OpenAI.

CoreWeave CRWV IPO 2025.03

$40 → ≈$82

The first GPU-cloud stock — double its offer price, but down by half from its $163 peak. The market has started telling stories apart from cash flow.

SpaceX (incl. xAI) IPO 2026.06

A $1.25T complex

Listed after absorbing xAI in February 2026; days after ringing the bell, it announced a $60B all-stock acquisition of AI coding tool Cursor.

Sources: CompaniesMarketCap 2026-07 · TechCrunch 2026-05-14 · StockAnalysis (CRWV, 2026-07-03) · Bloomberg/Fortune 2026-06-16

04 · Capital Markets · US · In the IPO queue

The biggest IPO in history is on its way

OpenAI S-1 filed confidentially

$852B

Valuation after its $122B raise in March 2026; in June it was revealed to have confidentially filed with the SEC, with a rumored IPO target of $1 trillion — timing may slip to 2027.

Anthropic S-1 filed confidentially

$965B

Post-money valuation after the $65B Series H in May 2026 (it was still $380B in February — 2.5× in a single quarter); filed its S-1 confidentially with the SEC on June 1, reportedly targeting a Nasdaq listing this October and raising over $60B. Annualized revenue above $44B.

SSI (Ilya Sutskever)

$32B

Founded by the OpenAI co-founder after his exit: no product, no revenue, $6B raised — the market's price tag on pure intellect.

Databricks

$134B

The data + AI platform; a fresh $5B raise, and a hot IPO candidate for 2026–27.

Sources: CNBC 2026-06-08 · Reuters 2026-06 · Anthropic official announcements (Series H 2026-05 · S-1 filing statement 2026-06-01) · PitchBook / Sacra

04 · Capital Markets · China · ByteDance special

ByteDance: the world's biggest unlisted tech company by revenue

$186B
2025 revenue (+20%) — about 7× OpenAI's; the world's largest unlisted tech company by revenue, and solidly profitable
36Kr/Sacra Jun 2026; overseas revenue topped 30% for the first time
≈$550B
Latest valuation (secondary-market share trades, Feb 2026) — second among all unlisted companies worldwide, behind only OpenAI
PE Insights Feb 2026; no timetable for a full IPO — the board has reportedly chosen to delay it
RMB 160B+
2026 AI capex budget (more than half of it for chips); per Bloomberg, internal talks of raising it toward $70B
Bloomberg 2026-05-27; a three-track chip strategy: NVIDIA + Huawei + in-house

Doubao app No.1 by users

MAU of 345M — China's top AI app (Qwen is second at 166M), with DAU past 100M; worldwide, second only to ChatGPT.

Doubao models × Volcano Engine No.1 in calls · No.1 in cloud

180 trillion tokens served per day — roughly 1,500× growth in two years; 49.5% of China's LLM public-cloud market, first place.

Jimeng AI · Seedance 2.0 World No.1

Topped the global blind-test board for video generation (Video Arena), beating Google Veo 3 and OpenAI Sora 2 — so hot at launch the waiting line hit 8 hours.

Coze

Agent-building platform, fully open-sourced under Apache 2.0 in July 2025 (9,000+ stars in 48 hours); serves tens of thousands of companies and millions of developers.

TRAE · AI coding

6 million global users within a year of launch, nearly 100 billion lines of code generated; 92% of ByteDance's own engineers use it.

Seed Edge · AGI ambitions

A long-horizon AGI research program launched in early 2025, with relaxed KPIs and patient funding; Zhang Yiming himself has returned to the front line, reading papers and interviewing top researchers.

The media's framing: "from app factory to AI factory" — Doubao only launched in August 2023, half a beat behind, yet in two years ByteDance took first place in users, token traffic, and cloud share through sheer spending power, internal horse races, and a full-stack layout (models–cloud–apps–chips). Note: TikTok's US business was handed to a US-majority joint venture in January 2026, with ByteDance keeping 19.9% — the company is also the most directly exposed player in the US-China standoff.

Sources: 36Kr/Sacra Jun 2026 · PE Insights Feb 2026 · Bloomberg 2026-05-27 · QuestMobile Q1 2026 · Volcano Engine FORCE conference 2026-06-23 (citing IDC) · Artificial Analysis Video Arena Feb 2026 · QbitAI Jul 2025 · GeekPark Dec 2025 · LatePost Jan 2025 · CNBC Jan 2026. Valuation/revenue figures are media or secondary-market estimates; ByteDance has never confirmed them.

04 · Capital Markets · China · One week, two bells — January 2026

The first LLM stocks land in Hong Kong

Zhipu · 2513.HK Listed 2026-01-08

Market cap ≈ HK$799.4B

"The world's first LLM stock." Day-one market cap of HK$57B, up roughly 13× in half a year (HK$1,793 per share, 2026-07-04). Now pursuing an "A+H" dual listing on the STAR Market, aiming to raise another RMB 15B. Its MaaS platform's annualized revenue grew 60× in 12 months.

MiniMax · 0100.HK Listed 2026-01-09

Market cap ≈ HK$108.7B

The public offering was oversubscribed 1,837×, breaking Kuaishou's record, in the biggest raise ever by an LLM company (HK$5.54B); on day one it became the world's first LLM company to touch a HK$100B market cap. Now HK$346.6 (2026-07-04) — up 110% from the offer price, but down nearly 40% from June's HK$553 peak.

A subplot: when lockups expired in February, both stocks whipsawed violently (Unisound swung even harder) — the market's enthusiasm is real, and so is the bubble risk. For investors and practitioners alike, volatility is part of this era.

Sources: Securities Times / QbitAI 2026-01-08/09 · Guancha 2026-01-09 · Sina Finance 2026-06-03 · The Paper Feb 2026 · Market caps/prices from 2026-07-04 market tracking

04 · Capital Markets · China's "Six Little Tigers"

Same starting line two years ago — no longer even the same race

CompanyStatusLatest valuation / market capFlagship productsDirection in 2026
ZhipuListed≈ HK$799.4B (2026-07-04)GLM series · MaaSFoundation models + B2B; sprinting for the "A+H" dual listing
MiniMaxListed≈ HK$108.7B (2026-07-04)M3 · Hailuo AI · TalkieAll modalities + a global consumer push
Moonshot AIFundraising champion$30B pre-money (up ~7× in half a year)Kimi · K2.6Staying the AGI course; preparing a Hong Kong IPO
StepFunRacing to IPOPre-IPO round near $2.5B, targeting a $10B valuationStep multimodal seriesBetting on "AI + devices": shipped on 42M+ OPPO/Honor flagship phones
BaichuanPivotedNo public valuation since 2025Baichuan-M3 medical LLMAll in on medical AI; out of the general foundation-model race
01.AIPivotedNo public valuation since 2025Yi series (mega-model training halted)Pivoting to enterprise AI deployment services

Sources: 36Kr, "After the Zhipu and MiniMax IPOs, the Six Little Tigers' paths diverge" · Caixin 2026-05-08 · Sina Finance 2026-06-09 · Caijing May 2026

04 · Capital Markets

The table is consolidating

Made it ashore

Zhipu, MiniMax

Secured a public-market funding channel — burning shareholders' money to keep training foundation models.

Still at the table

Moonshot AI, StepFun

One rides China's biggest funding hauls ($3.9B in six months); the other is bolted into the smartphone supply chain.

Switched lanes

Baichuan, 01.AI

Conceded that the foundation-model race is too expensive to stay in; pivoted deep into verticals like healthcare and enterprise services.

The lesson for you: foundation models are a game for giants and national champions — but application-layer opportunities are expanding exponentially. In the next chapter, we look at one of the application layer's defining products.
05

Agents: AI is no longer just chat

From "here's an answer" to "the task is done" — agents are the biggest paradigm shift of 2025–2026. Claude Code is our main case study, followed by two newer faces.

05 · Agents

Three leaps in three years: completion → conversation → agents

2021 · Completion

GitHub Copilot: you write half a line, it guesses the next one.

2022 · Conversation

ChatGPT: you ask, it answers — the copy-pasting is still on you.

2025 · Agents

Claude Code: give it a goal, and it does the work itself.

Claude Code's working loop in the terminal

① Understand

Reads the codebase and docs on its own to figure out the current state.

② Plan

Breaks the task down, lays out an approach, asks before any major action.

③ Execute

Writes code, edits files, runs commands, installs dependencies — many steps, back to back.

④ Verify

Runs the tests itself, reads the errors, and keeps fixing until they pass.

The key difference: when something breaks, it doesn't dump the error on you — it reads the error, fixes the code, and tries again. That is the line between a "tool" and a "colleague."
05 · Agents · Case study · Claude Code: the market votes with numbers

The fastest-growing category of developer tools ever

$2.5B+
Claude Code annualized revenue (Feb 2026): past $1B within 6 months of launch — Anthropic's fastest-growing product ever, with enterprise subscriptions making up over half
Anthropic official, Feb 2026
$60B
What SpaceX paid for AI coding tool Cursor (Jun 2026); its ARR went from $100M to $4B in 17 months
TechCrunch / Bloomberg 2026-06-16
100%
Multiple Anthropic and OpenAI engineers now say publicly that every line of code they commit is AI-generated; across Microsoft's repos it's ~20–30% (Nadella)
Fortune 2026-01-29 · LlamaCon 2025-04
Even the people who build AI no longer write code by hand — there is no harder proof that the tools have come of age.

Sources: Anthropic Series G announcement 2026-02-12 · Fortune 2026-01-29 · TechCrunch 2026-04/06

05 · Agents · Case study · Three typical Claude Code workflows (demoable live)

One sentence of instructions — how far can it run?

Case 1 · A website from scratch

"Build me a sign-up page for our class event — collect names, let people pick a session, export the list."

It will: scaffold the page → write the form logic → add data storage → run it locally for you to see. Something that used to take weeks of learning HTML/JS becomes a working version in 20 minutes.

Case 2 · Fixing a bug in an unfamiliar project

"This open-source project throws this error when I run it — fix it for me." (paste the error)

It will: read the error → trace the root cause through tens of thousands of lines → patch it → run the tests to confirm. Making sense of other people's messy code is exactly where AI saves you the most pain.

Case 3 · Automated data analysis

"Here's the survey spreadsheet — clean the data, run the cross-tabs, and generate a chart report."

It will: write the cleaning script → run the statistics → draw the charts → produce a one-page web report. No Python required — but you do need to judge whether the conclusions hold up.

The common thread: your role shifts from "operator" to the person who defines the ask, makes the judgment calls, and signs off on the results — the three skills the coming exercises will train.
05 · Agents · Case study · The "Lobster" (OpenClaw)

An open-source "Lobster" sets off an agent-raising craze

What it is Open source · Free

An all-purpose personal assistant that lives on your own computer: send it a message on WeChat / Telegram / WhatsApp and it organizes files, writes emails, fills in forms, crunches data, controls your devices — on call 24/7. Open-sourced by Austrian developer Peter Steinberger in November 2025; its mascot is a lobster 🦞, and the Chinese community calls deploying and tuning it "raising a lobster."

Why it took off

Skills as files: everything it can do is captured in SKILL.md documents you can accumulate and share — Fudan University even launched a "million-skill library." Model-agnostic: Claude, GPT, or DeepSeek can all serve as its brain; NetEase Youdao built a Chinese edition on top of it, LobsterAI.

380K+
GitHub stars — one of the fastest-growing open-source projects in history: 100K stars and 2 million visitors within a week of launch
GitHub check 2026-07-04: 381,646 stars
3 names
Clawdbot → Moltbot → OpenClaw: renamed twice in three months over Anthropic trademark complaints; in February 2026 the author was hired by OpenAI to build its agent-infrastructure team
Tencent News Feb 2026 · CNBC Mar 2026
The "lobster paradox" (Xinhua): it gets things done because its permissions run deep, and deep permissions mean risk — prompt-injection and skill-poisoning controversies keep coming; government agencies and banks restrict it even as tech parks nurture its ecosystem. The stronger the tool, the more you need to understand its boundaries.

Sources: GitHub API check 2026-07-04 · Wikipedia, "OpenClaw" · Xinhua 2026-03-23 · Tencent News 2026-02-02 · QbitAI Feb 2026

05 · Agents · Case study · Hermes Agent

Hermes: the agent that evolves itself

What it is MIT open source

A general-purpose personal agent released in February 2026 by Nous Research (the team behind the open-source Hermes model family), under the motto "The agent that grows with you." Deployed on your own server with data sovereignty front and center; one gateway connects 20+ chat platforms (WeChat / DingTalk / Feishu / Telegram…) and works with 200+ models; a beginner-friendly desktop app shipped just this June.

The big difference from the "Lobster" Self-evolving

The Lobster's skills have to be written by humans in advance; Hermes automatically distills every completed task into a skill document and keeps improving it with use, backed by persistent cross-session memory — the more you use it, the better it knows you. The media consensus: the two are "complements, not substitutes."

210K
GitHub stars in just 4 months — in April, 36Kr was still asking "is this the next Lobster?"; it has since sat at No.1 on GitHub's global trending list for weeks on end
GitHub check 2026-07-04: 208,890 stars
200+ / 20+
Models supported / chat platforms it plugs into — one agent, living inside every chat box you own
Official docs · featured on NVIDIA's official blog
The bigger picture: agents are now fiercely contested ground — on the commercial side, Manus (revenue above $100M) was acquired by Meta for over $2B, then halted by a Chinese security review; OpenAI poached the "Lobster's" author; ByteDance, Alibaba, and Tencent are all in. The open-source Lobster and Hermes are the share ordinary people can pick up for free, right now.

Sources: GitHub API check 2026-07-04 · 36Kr 2026-04-10 · Decrypt 2026-06-03 · Jiemian News 2025-12-31 · Guancha 2026-04-28

05 · Agent · Special topic · Agents move into the office (2026-06)

Coding assistants cross the line: giants race into finance & law

5M
Weekly active users of OpenAI Codex; about 20% are already non-developers (analysts, investors, bankers), growing 3x faster than developers
OpenAI official figures, 2026-06
110
Skills bundled in Codex's first 6 "role plugins" (plus 62 app integrations) — an entire job's workload packed into one plugin
OpenAI blog, 2026-06-02
+500%
Usage growth in the first 6 weeks after thousands of Freshfields lawyers deployed Claude for Legal
TechCrunch / Artificial Lawyer 2026-05

OpenAI · Codex role plugins 2026.06 follows

First six: data analysis, creative production, sales, product design, public equity investing (wired to Moody's / FactSet / LSEG institutional data), and investment banking (pitch decks, comps, due diligence); a legal plugin is "coming soon." Distribution is centrally managed by enterprise admins.

Anthropic · Claude industry editions 2025.07 first mover

Finance (2025-07: FactSet/S&P/PitchBook connectors, later an Excel sidebar and DCF-grade skills) → Legal (2026-05: 12 practice-area plugins) → open-source knowledge-work-plugins: ~15 role plugins for legal/finance/sales, free to install in Claude Code.

What this page is really saying: agents are turning from "a programmer's tool" into "a digital colleague for every white-collar role" — the entry-level work of banking analysts and lawyers is being packaged into plugins and sold. Bloomberg calls it a head-on fight between the two giants for the finance and legal markets. Your takeaway: the speed at which job skills get "plugin-ized" is the cash-out speed of that ranking on slide 4 — and Lecture 3 teaches you to flip it around: build your own "equity-investing plugin."

Sources: OpenAI blog 2026-06-02 · Bloomberg 2026-06-02 · Anthropic 2025-07-15 / 2026-05-12 · TechCrunch 2026-05-12 · GitHub anthropics/knowledge-work-plugins

05 · Agents

What this means for ordinary people

  • The coding barrier is gone; engineering judgment isn't. AI writes the code for you — it doesn't decide what to build, what good looks like, or whether it's right.
  • What gets replaced isn't "engineers" — it's people who can't use AI. Decomposing problems, specifying precisely, and verifying results are the new core competencies.
  • Question quality = output quality. Vague asks get vague results; spell out the goal, the constraints, and the acceptance criteria, and the output changes immediately.
  • Ordinary people carry almost no "experience deficit." This way of working only appeared in 2025 and only went mainstream in 2026 — you and a ten-year veteran are standing at nearly the same starting line. That almost never happens in history.
2026 · Year One of the Agent Era = a new starting line A ten-year engineer Ten years on the old track You, today The AI era

Miles logged on the old track don't carry over to the new one — the road ahead is the same length for both.

06

AGI & Embodied AI: When Intelligence Gets a Body

Chat and code were only the first half. When large models climb into robot bodies, the revolution steps off the screen and into the physical world.

06 · Embodied AI · The AGI countdown

How far is AGI? The lab chiefs agree: years, not decades

Demis Hassabis · DeepMind 2026.05

"About 3–5 years to AGI — around 2030." Its impact will be "roughly 10x the Industrial Revolution, happening 10x faster — unfolding over a decade, not a century."

Sam Altman · OpenAI 2025–2026

"We are now confident we know how to build AGI." He also predicts robots doing real-world physical work by 2027 — exactly this chapter's subject.

Dario Amodei · Anthropic 2026.01

Avoids the term AGI, but expects "powerful AI" as early as late 2026–early 2027 — and warns, in a 38-page essay, of civilization-scale risks that come with it.

Moravec's Paradox (1988): "It is comparatively easy to make computers exhibit adult-level performance at chess, and difficult or impossible to give them the perception and mobility skills of a one-year-old." Giving intelligence a body is the last — and hardest — step. That is the battleground of embodied AI.

Sources: Axios 2026-05-26 · OpenAI blog "The Gentle Singularity" 2025-06 · Forbes 2026-01-28 · Moravec, "Mind Children," 1988. Note: industry timelines skew aggressive (Amodei's "90% of code by AI in 6 months" ran late) — watch what they build, not just what they say.

06 · Embodied AI · The technical inflection point

Large models climb into robot bodies: VLA models

VLA (Vision-Language-Action) = the LLM of robotics: see the scene, understand plain language, output actions directly — for the first time, the robot "brain" problem has a general solution.

Google · Gemini Robotics 1.5 2025.09

"Think before acting"; skills transfer zero-shot across robot bodies — one robot learns, every robot knows.

Figure · Helix 02 2026.01

One model controls walking, balance, and two-hand manipulation; in the demo it loaded a dishwasher fully autonomously for 4 minutes, no human in the loop.

Physical Intelligence · π0 Valuation 2x in 4 months

Folds laundry, assembles, sorts; valuation jumped from $5.6B to a reported $11B+ — capital's price tag on a "general-purpose robot brain."

The timeline: in 2023, Google's RT-2 first proved that web-scale knowledge can become robot actions; in 2025–2026, three approaches closed the "see–understand–act" loop at once. That is why embodied AI is exploding now, not a decade ago.

Sources: Google DeepMind blog 2025-09 · Figure 2026-01-27 · TechCrunch 2026-03-27

06 · Embodied AI · The contenders (mid-2026)

Humanoids: America races on brains, China races on shipments

Tesla Optimus V3 production this summer

The Fremont line is cleared for a summer-2026 production start; Musk himself tempers expectations — "production will be excruciatingly slow at first," and today's few hundred units are mostly "learning."

Figure Valued at $39B

Figure 02 worked a BMW plant for 11 months: loaded 90,000+ parts across 30,000 vehicles; Figure 03 targets mass production and the home.

Boston Dynamics · 1X

Production electric Atlas (CES 2026) is sold out — Hyundai (committed to deploying 25,000 units) and DeepMind took the entire run; 1X NEO home humanoid ships in the US at $20K or $499/month.

Unitree IPO within weeks

5,500+ humanoids shipped (claims world No. 1); G1 starts at RMB 99K; STAR Market registration effective after a record 73-day review — "the first humanoid-robot stock."

AgiBot 5,100+ shipped

Co-founded by Zhihui Jun; third-party reports rank it world No. 1 by annual shipments (a different tally than Unitree's — both claims coexist). Target for 2026: tens of thousands.

UBTech · Galbot

UBTech (9880.HK) at about HK$49.3B market cap (2026-07-06), Walker S2 training on BYD and Zeekr factory lines; Galbot, backed by the national AI fund, runs 100+ unmanned pharmacies.

Two snapshots of China speed: at the 2026 Spring Festival Gala, Unitree's G1+H2 staged the world's first fully autonomous humanoid martial-arts swarm; in April, the winner of Beijing's humanoid half-marathon clocked 50:26 — about 7 minutes faster than the human world record. Underneath it all, the supply chain: 75%+ of core components are domestically sourced (CAICT; Morgan Stanley puts it above 90%), and China shipped 14,400 humanoids in 2025 — 84.7% of the world's total.

Sources: SSE filings 2026-03/07 · Xinhua 2026-01-08 / 2026-04-19 · Hyundai CES 2026 · Teslarati/Electrek 2026 · UBTech market cap from portfolio tracker, 2026-07-06

06 · Embodied AI · Dexterous hands & the trillion-dollar horizon

The hardest part is a hand; the biggest prize is an era

Why hands are hardest

The human hand has about 27 degrees of freedom, combining strength with millimeter precision — Moravec's Paradox at its sharpest: a robot can backflip, yet may not tie a shoelace.

The DoF race

Shadow Hand, 24 joints (the academic benchmark) → Tesla Optimus Gen3, 22 DoF per hand (actuators moved to the forearm, tendon-driven) → China's LinkerHand tops out at 42 DoF, with RMB-20K-class models crashing the price. Touch is catching up too: Meta's Digit 360 fingertip senses 18 modalities at human-level acuity.

The dexterous-hand market

About $5B by 2030, growing 64.6% a year (QYResearch; estimates vary widely across firms).

Wall Street's disagreement is itself the signal

Morgan Stanley: $5T and 1 billion units by 2050; Citi: $7T and 650M units; Goldman the conservative: just $38B by 2035 — two orders of magnitude apart, yet not one of them says "this won't happen."

Musk's wager

About 80% of Tesla's future value will come from Optimus (2025-09); at Davos 2026 he doubled down: "robots will outnumber humans within 3–5 years."

China's position

"Embodied AI" entered the Government Work Report in 2025 and was named again in 2026, and the 15th Five-Year Plan lists it as an industry of the future — policy, supply chain, and shipments all aligned: China's strongest relative bet in AI.

Your takeaway: embodied AI = a decade-long compounding track of AI × mechanics × materials × supply chain. For the open-topic exercise on slide 41, pick an angle here — it's the best possible starting point.

Sources: Morgan Stanley 2025-04 · Citi GPS · Goldman Sachs 2025-10 · CNBC 2025-09-02 · TMTPOST 2026-01 · Xinhua 2026-03-09 · QYResearch

07

Hands-on exercises

Five exercises, all driven by the same framework:

Goal

State in one sentence what you're going to produce

Tool

Which AI to use, and how to get started

Deliverable

What you hand over for review when you're done

Review

Which prompt mattered most? Where did the AI go wrong?

Exercise 1 · Data infrastructure

Have an agent build a "scrape → clean → store" data pipeline (market data / news / weather — pick one) that runs repeatably.
Deliverable: a working script + one clean data file.

Exercise 2 · Investment analysis

Pick an AI company from today's talk (Zhipu / MiniMax / NVIDIA…), have the AI verify the data online, then produce a one-page investment brief: business, valuation, risks.
Deliverable: a shareable one-page web page or document.

Exercise 3 · Video transcription

Hand the AI a talk or lecture video: transcribe it → turn it into structured notes with subheadings and standout quotes.
Deliverable: notes polished enough to publish as-is.

Exercise 4 · The Hermes tool

Deploy Hermes Agent on your own computer, connect it to a chat app, direct it through one task, and let it distill the task into a skill.
Deliverable: a screenshot of the successful conversation + the resulting skill document.

Exercise 5 · Your own pick

Use any tool you learned today to solve a real problem from your own study or life. The one requirement: state the goal and acceptance criteria before you start, and afterwards review where the AI went wrong.

A hint: these exercises don't reward whoever finishes fastest, but whoever states it most clearly and verifies it most rigorously — question quality = output quality.

08

The Pro Workflow: Lead AI Like You'd Lead a Team

You can do the exercises — now how do you turn one into a real project? A complete playbook from preparation through iteration, review, and operations. One person plus a fleet of agents is an engineering team.

08 · Pro workflow · Before you start

Write three documents first: the AI's "onboarding kit"

① Knowledge base Sharpen the axe

Sweep the web with intent: papers, research reports, official docs, competitors, the latest developments — have AI do the deep research and distill it into knowledge-base.md. Every new session reads it first, and the AI is instantly "in the field."

② Data infrastructure The foundation

Where the data comes from, how it's fetched, cleaned, and stored — get the pipeline running first (that's Exercise 1), then add a "data dictionary" explaining every field and its definition.

③ Goals & constraints brief The contract

One page stating the goal, acceptance criteria, boundaries (what you won't do), and technical constraints. It becomes the first version of the project's CLAUDE.md — when the AI drifts, this is what you align it against.

Rule of thumb: these three documents cost you 20% of the time and save 80% of the rework — the quality of AI output is bounded by the quality of the context you hand it.
08 · Pro workflow · While the project runs

Small steps, tight reins

① Code with tests

Have AI generate each module together with its unit tests — tests are the reins on AI iteration

② Run to green

A step counts as done only when every test passes; if it's red, keep going

③ git commit

Checkpoint every small win; always commit before letting AI make big changes

④ Update CLAUDE.md

Hit a pitfall, set a rule — write it down immediately: the AI's "mistake journal"

Why tests are the reins

AI edits code fast and boldly; unit tests are the only thing that automatically tells you something broke. A project without tests is one AI fears to touch — and you can't sleep on.

Why CLAUDE.md must stay current

Every new session starts with amnesia; CLAUDE.md is the AI's only long-term memory. An unwritten rule = the same pit stepped in N times. (This very slide project's CLAUDE.md grew exactly this way.)

git, commits, and CLAUDE.md are all taught in Lecture 2 — what this page adds is stringing them into a rhythm.
08 · Pro workflow · Review & fixes

Never let the AI grade its own homework

A second pair of eyes

Separate the session that writes code from the one that reviews it: open a fresh session (or a dedicated code-review command) and let an AI with no stake pick it apart — correctness, security, performance.

Humans review the critical paths

Anything touching money, data deletion, or public releases gets human eyes, no exceptions — AI covers the breadth, you cover what's vital.

Bug fixing in four steps

Paste the full error text to the AI → have it explain the cause before touching anything → fix → run the regression tests. Then add a test for it, so this bug can never come back.

The mindset: treat AI output as work submitted by "a brilliant but overconfident intern" — welcome the enthusiasm, inspect it strictly.
08 · Pro workflow · Shipping & daily operations

Make the project take care of itself

Layered tests, then ship

Unit → integration → end-to-end, all green before launch; a pre-launch checklist: rollback plan, data backup, monitoring in place. Deployment goes in a script — the human just presses Enter.

Logs are the project's EKG

Log consistently; watch live with tail -f. Then go further: have AI read the logs on a schedule — "every morning, summarize last night's errors and suggest fixes."

Autonomous daily operations

Scheduled jobs (cron / Hermes) put an agent on duty: health checks, backups, data refreshes, an auto-generated weekly report — the project goes from "you feeding it" to "it growing on its own."

The loop closes: problems found in operations flow back as input to the next iteration, while the knowledge base and CLAUDE.md keep growing. Once this wheel turns, one person plus a fleet of agents is a complete engineering team.

One last slide

The best way to predict the future
is to build it.

The curve is still early. Now open a terminal and start Exercise 1.

GitHub