A Personal Perspective · Mid-2026

AI in 2026

Where we are, how we got here, and what's next

https://ai-in-2026.xiaomum.com

Agenda

Four parts, one thread

Part 1 From my own desk Two real stories — no theory, no benchmarks.
Part 2 How we got here A twelve-month industry timeline.
Part 3 Under the hood RL, distillation, and the agent loop.
Part 4 What's next My takes, and open questions.
01

From my own desk

Two real stories from the first half of 2026

No theory, no benchmarks — just things that actually happened to me.

Story 1 · The problem

200 GB of GoPro footage, gone

  • May Day trip, ~200 GB of GoPro video on a microSD card
  • Pulled the card without ejecting safely → card became unreadable
  • Every free recovery tool I could find: failed
  • Paid "pro" recovery software: only partial recovery — glitched video, or audio gone entirely

Personal experience, May 2026

Story 1 · The fix

Claude Code fixed it, byte by byte

$ hexdump -C /dev/sdb | less # first: look at the raw disk before touching it
$ photorec · exiftool · ffmpeg # forensic CLI tools, chained
$ verify: every clip, video + audio streams
✔ 214 clips recovered — 100% intact

Pointed Claude Code at the raw block device. It diagnosed first, then acted.

▶ Live animation — Byte Surgeon

Personal experience, May 2026

Story 2 · The setup

My dual-host desk setup

Work PC Home PC
▼
USB Switch
▼
USB Hub (DC jack, never used)
▼
Keyboard · Mouse Speaker · Mic ⚠

Full-tunnel VPN forced two hosts; one button to switch everything — almost perfect.

[ photo: speaker + USB hub — save as images/usb-hub-setup.jpg ]

Story 2 · The diagnosis

A gremlin, diagnosed in 60 seconds

symptom: speaker & mic randomly silent; kbd/mouse on the same hub: rock solid
PS> Get-PnpDevice | ? FriendlyName -match 'USB|Audio' # opencode writes PowerShell
→ every device: Status OK — hardware is healthy
✔ hypothesis: insufficient USB power. Verify: plug audio into the motherboard.

The hub had a DC power jack I'd never used. Plugged it in → fixed, instantly.

▶ Live animation — Sixty Seconds

Personal experience, July 2026

The pattern

What the two stories share

Not answers — actions Both were solved by an agent that acted: ran tools, inspected raw state, tested a hypothesis.
Beyond consumer software GoPro: binary-level forensics no recovery app could match.
Like a real engineer USB hub: systematic diagnosis in ~60 seconds — not a guess.

My take: in 2026, the real shift isn't "smarter chat" — it's that AI finally has hands.

02

How we got here

A twelve-month timeline: H2 2025 → mid-2026

Facts, not takes — every source is in references.md.

Timeline · Models

The frontier model war, month by month

Aug 2025 OpenAI ships GPT-5 — chat and reasoning, unified1
Nov 2025 Three frontier launches in two weeks: GPT-5.1 · Gemini 3 Pro · Claude Opus 4.52
Dec 2025 OpenAI's reported "code red" over Gemini 3 → GPT-5.23
2026 Monthly cadence: Opus 4.6→4.8 → Fable 5, GPT-5.3→5.6, DeepSeek V44
1OpenAI, "Introducing GPT-5", 2025-08-07 2OpenAI GPT-5.1 (11-12) · Google Gemini 3 Pro (11-18) · Anthropic Opus 4.5 (11-24), 2025 3Magica/Galaxy AI model tracker; Sina Tech on the reported "code red", Dec 2025 4PromptZone AI model release tracker, verified 2026-07-16 Full URLs: references.md § Frontier releases

Timeline · Open weights

Open weights caught up

−50% DeepSeek V3.2-Exp API price (Sep 2025), via sparse attention1
3×🥇 IMO · IOI · ICPC gold level, DeepSeek V3.2-Speciale (Dec 2025)2
6× cheaper than GPT-5.5 — GLM-5.2 (Jun 2026), MIT-licensed 753B; vendor coding claims3
2.8T parameters — Kimi K3 (Jul 2026), largest open-weights model ever; weights drop Jul 274

My take: open models now trail the frontier by months, not years — and they set the price floor.

1DeepSeek V3.2-Exp launch, via PromptZone tracker, 2025-09-29 2DeepSeek official announcement (V3.2 & Speciale), 2025-12-01 — vendor-reported 3OpsMatters, "GLM-5.2 Review", 2026-07-08 — headline benchmarks are Zhipu's own 4Xinhua, "中国企业发布全球最大规模的开源模型 Kimi K3", 2026-07-17 Full URLs: references.md § Open weights

Timeline · Agents

Agents became the product

2025 Chatbots → systems that act Browse, code, operate computers. Claude Code / Skills go mainstream.1
Dec 2025 Meta acquires Manus Multi-billion-dollar bet on general-purpose agents.2
Jan 2026 OpenClaw surge Open-source agent framework hits GitHub's global top 10 in ~10 days.3
1Anthropic, Claude Skills launch, Oct 2025 2Tencent News, 2025-12-30 — price not officially disclosed 3AI Skill Navigation timeline — single third-party source, approximate Full URLs: references.md § Agents

Timeline · Money & jobs

Money, compute — and jobs

OpenAI · round closed Mar 20261 $852B
Anthropic · May 2026, IPO filed in June2 ~$965B
  • Both labs started designing their own AI chips3
  • AI cited as the #1 reason for US tech layoffs, spring 20264
  • Chinese labs joined the market: Zhipu (Z.ai) listed in Hong Kong (Jan 2026); Moonshot reportedly prepping a HK IPO5
1kingy.ai funding tracker (company announcements aggregated), 2026-03-31 2Launch Consulting; OSAS AI on the IPO filing, May–Jun 2026 — secondary reporting 3Crypto Briefing, 2026-07-14 4AIToolsRecap, citing Challenger, Gray & Christmas, 2026-05-11 5Sina Finance, 2026-01-08; Moonshot IPO is reported, not filed Full URLs: references.md § Industry & capital

Timeline · Regulation

June 2026 — the first model recall in history

Jun 9 Anthropic launches Claude Fable 5 — first "Mythos-class" model, above Opus1
Jun 12 US export-control order: no foreign access — Fable 5 & Mythos 5 go 404 worldwide2
Jul 1 Commerce Dept lifts the order; access restored after ~3 weeks3
Jun 26 → Jul 9 OpenAI's GPT-5.6 clears a government security review first — the first gated frontier launch4

My take: frontier capability is now a national-security matter — expect more governance, not less.

1Codersera launch guide; ccleaks — Anthropic press release, 2026-06-09 2Huxiu, "美国首次以出口管制强制下线 Anthropic 最强 AI 模型 Fable 5", 2026-06-15 3Yicai via MOFCOM export-control portal, 2026-07-03 4o-mega.ai, citing TechCrunch & Engadget, 2026-07-08 Full URLs: references.md § Regulation
03

Under the hood

How are these capabilities actually made?

Three ideas: post-training, RL, and the agent loop.

Concept 1 · Post-training

The post-training pipeline

1 Pre-training predict the next token
→ knowledge, no manners
→
2 SFT imitate demonstrations1
→ follows instructions
→
3 RL optimize rewards2
→ prefers good answers,
discovers reasoning

My take: the 2025–2026 capability jump came mostly from the RL stage, not from bigger pre-training.

1Ouyang et al., "Training language models to follow instructions with human feedback" (InstructGPT), arXiv:2203.02155 2DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL", arXiv:2501.12948 Full URLs: references.md § Training methods

Concept 2 · Reinforcement learning

Three flavors of RL for LLMs

RLHF1 A reward model trained on human preferences scores the answers. InstructGPT lineage.
RLVR2 Verifiable rewards — math answers and code tests checked by rules. DeepSeek-R1 lineage.
DPO3 Skips RL entirely; optimizes preference data directly. Cheap, but no online exploration.

RLVR is why models learned to think longer: trial and error discovered chain-of-thought on its own.

1Ouyang et al. 2022 (InstructGPT), arXiv:2203.02155 2DeepSeek-AI 2025 (DeepSeek-R1), arXiv:2501.12948 3Rafailov et al. 2023 (DPO), arXiv:2305.18290 Full URLs: references.md § Training methods

Concept 2.5 · A common confusion

Distillation is not RL

Distillation A small student imitates a big teacher's outputs. Plain supervised learning.
Copying the top student's homework.
≠
Reinforcement learning The model samples its own answers and optimizes for reward.
Doing your own problem sets, graded pass/fail.

DeepSeek-R1 was trained mainly with RL — then its reasoning was distilled into Qwen / Llama small models.1
My take: distillation spreads capability; RL is what creates it.

1DeepSeek-AI 2025 (DeepSeek-R1, incl. distilled variants), arXiv:2501.12948 Full URLs: references.md § Training methods

Concept 3 · Agents

An agent = model + tools + loop

Observe → Think (LLM) → Call a tool → Observe result

↺ repeat until done — tools: shell · file system · browser · APIs (MCP)12

  • Tool use is trained (RL on real tool-call traces)
  • …and scaffolded (harnesses like Claude Code)

My take: "hands" appeared because long-horizon reasoning and reliable tool use crossed the usability threshold together.

1Yao et al. 2022 (ReAct: reasoning + acting loop), arXiv:2210.03629 2Anthropic, Model Context Protocol spec, Nov 2024 Full URLs: references.md § Agents

Case study · The training ground

The environment factory

1 Generate synthesize a task
(a repo + a failing test,
a broken server, …)
→
2 Execute spin up a container;
the agent works inside it
for many turns3
→
3 Verify run the test suite →
programmatic reward →
GRPO update
$10M+ reported yearly spend on RL environments — Anthropic; OpenAI signs seven-figure env contracts1
5,000+ isolated environments launched within 10 seconds — MiniMax; tens of thousands concurrent1
80–90% of RL wall-clock time is rollouts — the environment is the bottleneck12

My take: hands are taught in gyms. The environment became the data engine — and the moat.

1guanghan.ai, "Enter the Scaling of RL Environments", 2026-04-09 — reported figures, not official 2LessWrong, "…one more RL scale up…", 2025-09-03 — environment quality as the bottleneck 3Sapphire Ventures, "Reinforcement Learning: Learning by Doing", 2026-04-13 Full URLs: references.md § Training methods

Case study · The Fable 5 files

How is a Mythos-class model trained? The clues

Trained inside graded environments Fable 5 learned to notice when it's being graded — and the system card traces that to exploitable graders in its own RL training environments.1
Reasoning is trained, never dictated An accidental chain-of-thought supervision bug during training was disclosed as a safety incident — readable CoT is a monitored asset.1
Post-training > pre-training Cursor disclosed a 20× RL scale-up where post-training compute passed pre-training; Anthropic credits its 2026 gains to CAI + RL.3
Safety is a wrapper Alignment training (constitution), then deployment-time probes, classifiers, and fallback routing — even silent PEFT downgrades on ~0.03% of traffic.12

My take: the recipe is public — the task distributions and reward signals are the real moat.

1Anthropic, "Claude Fable 5 & Claude Mythos 5 System Card" (264 pp.), 2026-06-09 2Shivam (cvam.sight), system-card read-through, Jun 2026 — third-party summary 3Digital Applied, "The Post-Training Revolution: RL Is the New Moat", 2026-05-17 Full URLs: references.md § Training methods

Full circle

Back to my two stories

GoPro recovery The agent loop over forensic tools: hex-dump → hypothesize → run tool → verify.
USB diagnosis Systematic hypothesis testing, like an engineer — not a lucky guess.
Not magic Trained behavior + tool access.

The shift: from "AI gives answers" to "AI runs the process".

04

What's next

My takes, and questions I can't answer

My takes · Opinion

Three takes for engineers

01 Value shifts from writing code to designing systems + verifying AI output
02 Review skills, taste, and problem decomposition become the real leverage
03 Agents keep absorbing "mid-level implementation" work — labs now sell into consulting / enterprise delivery1
1nerdleveltech on OpenAI's & Anthropic's enterprise ventures, 2026-05-04 — takes 01 & 02 are opinion Full URLs: references.md § Industry & capital

Open questions

Questions I can't answer

Verification at scale How do we verify agent work beyond "it looks right"?
The talent pipeline If juniors delegate implementation on day one, where do seniors come from?
The curve Does capability flatten — or does the monthly cadence just keep going?

Discussion welcome — find me after the talk.

Fin

References

All sources: references.md (in the GitHub repo)

https://ai-in-2026.xiaomum.com