AI Dev 26 x SF: Andrew Ng: The Future of Software Engineering

At AI Dev 26 x San Francisco, Andrew Ng discussed the rapid evolution of software engineering driven by AI coding agents and introduced new tools to support this shift:

- The Shift in Software Development - New Bottlenecks and Generalist Skills - Job Market Perspective

Andrew also announced:

Context Hub: A tool designed to provide AI agents with up-to-date documentation to prevent hallucinations and the use of deprecated APIs.

Code Dream: An interactive learning environment featuring AI-driven video conversations and a browser-based terminal for practicing with modern coding agents.

Ng / DeepLearning.AI — The Future of Software Engineering (AI Dev 26 x SF)

Andrew Ng keynote at AI Dev 26 x San Francisco, published on the DeepLearningAI YouTube channel — 20 May 2026, ~19 minutes. The wiki’s first solo-headlining Andrew Ng anchor (his prior wiki appearance was a brief cross-reference in MIT March 2026 via the “we need to unbig in AI” quote). The talk is a tightly-structured 19-minute future-of-software-engineering address with two product announcements at the end: Context Hub (for AI agents) and Code Dream / Code Realm (for humans) — both framed as parallel skill development tools.

TL;DR

• “My own coding is pretty much 100% AI” (~2:48) — Ng’s load-bearing personal claim. “Many of the frontier teams are trending toward 100% almost 100% of code written by AI. And this creates a real acceleration that getting only 80% of the way there feels very different from.” Not religious (“if you’re working for NASA and you need to write 50 lines of code for a spaceship, go ahead, write it by hand”) but observable in cohort behaviour at the frontier-team altitude.
• The PM bottleneck thesis (~4:20–6:00, the talk’s load-bearing rhetorical centerpiece). As AI coding accelerates programming 10–100×, deciding-what-to-build becomes the new bottleneck. Empirical signal: PM-to-engineer ratios shifting from a typical 1:8 (one PM keeping eight engineers busy) toward 1:2, then 1:1, then the collapse into single-human-doing-both — “the only thing that can move even faster is you take those two people and collapse them into a single human. Engineers who shape products or product managers that can code can move really fast.” The wiki’s clearest AI-conference-altitude articulation of the engineer + PM → single AI-native-builder collapse.
• Cascading bottlenecks (~6:30–9:30, the talk’s most operationally-useful extension). The PM bottleneck is the first wave; design, legal/compliance, marketing, and sales bottlenecks follow. “If you spend 3 months writing code and legal takes a month to sign off, it’s okay, but if you take a day writing code, you’re going to wait a month for legal, it’s like, boy.” The implication: AI-native teams need to be small teams of generalists where each person can do some-of-all functions — “if a team needs software, product, design, legal, and marketing, and it’s a team of two, by definition these two people have to have some skills in all of these functional areas.”
• The AI-job-apocalypse counter-claim (~9:40–11:05). “I just don’t see the AI job apocalypse happening anytime soon… my teams, we can’t find enough of these people” (referring to AI engineers with the building-blocks-knowledge + generalist-skills mix). Cites recent business-media coverage acknowledging “job apocalypse being delayed” + Federal Reserve Bank of Philadelphia study. Ng’s framing is calibrated, not Pollyanna-ish: “I know a lot of people are feeling job insecurity at all levels of [seniority]… but the job apocalypse is plausible to talk in some circles, I just don’t see it happening.”
• The AI-engineer hiring rubric (~6:27–7:25). Three-criterion checklist Ng’s teams use when hiring: (1) ability to use coding agents effectively (Claude Code, Gemini, Codex, Open Code, or something else); (2) robust knowledge of the proliferating building blocks — “a challenge because there’s so many of them. So many and they change so quickly”; (3) generalist skills — basic product management or other functional skills. “A combination of skills on the ability to how to build things as well as the ability to know what to build.”
• The building-blocks lego-bricks metaphor (returns throughout, named in opening at ~0:43): software components — both AI (LLMs, RAG, agentic workflows) and non-AI (UI, persistence, identity/auth) — “proliferating at a speed like we’ve never seen” under AI coding. The combinatorial growth in possible compositions is what AI coding agents enable and what human-skill development must keep pace with.
• Parallel skill development (~11:07–12:11): the framing under which Ng pitches both product announcements. “As our coding agents become more capable, our people need the complimentary skills to help us to drive the coding agents in the appropriate way.” Two tracks: agents getting more capable (Anthropic agent skills, etc.) + humans developing complementary skills.
• Context Hub announcement (~12:12–14:16): a tool Ng built with Vivek Prasad and Sanyam Hota that gives AI agents most up-to-date documentation. Solves the knowledge-cutoff-leads-to-deprecated-API-calls failure mode. Concrete example: if you ask Claude Code today to call the OpenAI API, it’ll use the older deprecated chat completions API because that dominates the training data; Context Hub returns the latest 600-line markdown docs so the agent knows about GPT-5.5 and the newer responses API. CLI: chob search OpenAI (lists packages), chob get OpenAI/chat (fetches latest docs).
• Code Dream / Code Realm announcement (~14:18–17:50, the talk’s first-public announcement): not a course — a conversation. Video-call interface where the user can talk to Ng (or an AI version of Ng) in real time, paired with a browser-based terminal for hands-on practice (npm install, run open code, paste commands from the slide). Live demo of the “Codoji build environment” (ASR likely garbling the product name — “Code Dream” / “Code Realm” used interchangeably on stage). Available in preview as of the talk date.

What was actually ingested

Full ~19:21 transcript via the yt-dlp fallback path (Playwright path failed at 60s timeout, then yt-dlp --write-subs --sub-langs en --sub-format vtt succeeded with 502 deduped ASR segments). The same hour-aware VTT parser developed during the 22 May Caldwell ingest was reused without modification. Names noted with ASR-uncertainty: “chob” (the CLI binary name for Context Hub — likely correct), “Codoji” and “Code GG” (likely ASR mangling of Code Dream / Code Realm / Code Realm Build — Ng uses Code Dream and Code Realm both on stage; the product-page-canonical name is not in the transcript). Stage cues and audience reactions retained. Full raw at raw/videos/2026-05-20-ng-deeplearningai-ai-dev-26-sf-future-of-software-engineering.md.

Substantive content

1. The building-blocks-proliferation framing (~0:43–2:12)

Ng opens with the Lego-bricks metaphor he used at the first AI Dev a year earlier: software is assembled from building-block components — tools, frameworks, APIs — and “as the number of disparate building blocks you have grows, the way you can combine them to create interesting software grows exponentially or grows combinatorially.” The 2026 development: “over the last year, the rise of AI coding agents has made it far more effective for all of us to quickly put these building blocks together.”

Two classes of building blocks:

• AI building blocks: large language models, RAG, agentic workflows.
• Non-AI building blocks: UI components, persistence/databases, identity/auth managers.

“These building blocks are proliferating at a speed like we’ve never seen. There are more and more building blocks every single day, both AI and non-AI, for all of us to use.” This is the substrate observation under everything else in the talk — the building-blocks-proliferation rate sets the pace for what coding agents can compose and what humans need to know about.

2. The 100%-AI-coding claim (~2:12–4:18)

Ng poses the question many teams debate — “how much should we use AI coding agents?” — then names his own position: “my own coding is not 80% AI, it’s actually pretty much 100% AI. And I want to share with you why I think this even last gap makes a big difference.”

The argument: when 20% of code is human-written, the human review of that 20% becomes the bottleneck. “AI coding is fast, human review of that 20% of code just takes forever. Whereas a team that’s 100% AI coding just gets a lot more done in a short time.” Ng extends this to code review more broadly: “I know that a lot of teams that still have AI write code, but have a human review every single line of code. But for myself and a lot of my teams, we find that if I got to review the code, then I become the bottleneck, and it also doesn’t work.”

Calibrating caveat: “This is not a religion where we should never be allowed to write code by hand. But what I’m seeing is many of the frontier teams are trending toward 100% almost 100% of code written by AI. And this creates a real acceleration that getting only 80% of the way there feels very different from.”

This is the wiki’s clearest AI-conference-altitude statement of the unconditional-frontier-team-100%-AI-coding norm. Productive tension with Momentic’s “Codex only makes you a 10× engineer if you weren’t a 10× engineer to begin with” (conditional-on-baseline) and structural convergence with CS153’s unconditional 1,000× engineer framing (published the same day).

3. The PM bottleneck — the talk’s load-bearing rhetorical centerpiece (~4:20–6:00)

Ng’s “product management bottleneck” observation, first written in the deeplearning.ai batch newsletter the prior July: “When I’m trying to build stuff, you’ll often have an idea, build the prototype — that’s software engineering — and then go to users, get feedback — that’s PM or product management work — use those ideas to come and refine the software.”

“AI coding has made that 10 or 100 times faster. And so this means that deciding what to build or the product management work becomes the new bottleneck rather than the actual building.”

The empirical signal — PM-to-engineer ratios collapsing:

• “We used to have product manager to engineering ratios of, you know, one PM would keep eight engineers busy. These ratios typically vary in a small range, one to eight PM to engineer, one to seven common in many companies.”
• “But the ratio started to trend toward, you know, one to two was really weird, and then one to one, which is even more weird.”
• The structural endpoint: “What I’m seeing is rather than one engineer and one PM, the only thing that can move even faster is you take those two people and collapse them into a single human. Engineers who shape products or product managers that can code can move really fast. And that speed is a huge advantage.”

The team-shape consequence: “For more and more AI natives, the trend is that very small teams, they’re all kind of generalists where the engineers can do some PM work, the PMs can do some engineering work, and everyone kind of knows almost everything, and the small team can move really fast.”

This maps directly onto AI-founder-type three-role org structure published the same day — Hu’s prescription is everyone ships, including non-technical staff; Ng’s empirical observation is the cohort-behaviour signal under that prescription.

4. Cascading bottlenecks (~6:30–9:30)

The PM bottleneck is the first wave; Ng catalogues the second-wave bottlenecks that emerge once engineer-plus-PM-collapse-into-one clears the first one:

• Design bottleneck — “you design something and she implements it or even better, maybe the designer just implements it in code without going through a design tool like, you know, Figma, and that even speeds it up.”
• Legal / compliance bottleneck — “one thing that used to be painful, but is even more painful now, is the legal compliance bottleneck. If you spend 3 months writing code and legal takes a month to sign off, it’s okay, but if you take a day writing code, you’re going to wait a month for legal, it’s like, boy.”
• Marketing bottleneck — “For many of our teams, we write code and ship products so fast that marketing has a hard time scrambling to keep up with what these engineers are doing to figure out how to tell people about it.”
• Sales bottleneck — named in passing.

The doctrinal answer: small teams of generalists where each person can do some-of-all-functions. “It turns out that an engineer that’s not a marketer, but that knows how to prompt AI well — maybe they’re not an amazing marketer, AI actually makes really weird marketing decisions, but, you know, they could do some basic marketing work.”

The math constraint that forces generalism: “If a team needs software, product, design, legal, and marketing, and it’s a team of two, by definition these two people have to have some skills in all of these functional areas.” Tip on legal: “I often have AI deal with legal stuff as a first draft and then take it to a real lawyer to sign off before I launch something.” (Not advising anyone to be a lawyer, but the AI-first-draft → human-sign-off pipeline is operational.)

Brief scaling note Ng doesn’t have time to develop: “How to structure multiple AI native teams to work together — because not everything can be done by a 10-person team, but if you have a 100-person team of many AI native teams with limited communication and clear API boundaries, that also helps to scale beyond a single team.” Conway’s-law-for-AI-native-orgs as the future-research-direction placeholder.

5. The AI job apocalypse counter-claim (~9:40–11:05)

Ng’s load-bearing position on the will-AI-take-jobs question, calibrated rather than dismissive:

“In some Silicon Valley circles, it’s trendy to talk about the job apocalypse — this idea that AI would take all jobs almost all jobs and the rewriting in this phase. Frankly, I’m not really seeing that.”

The empirical anchor: hiring demand for the AI-engineer-with-generalist-skills profile. “Massive unmet demand for a lot more engineers with anywhere near this mix of skills. Frankly, even my teams, we can’t find enough of these people.”

The cross-citation Ng uses — business-media converging on the job-apocalypse-delayed framing: “I’ve been encouraged that over the last month or two a lot of the business media has been getting the story right that this AI job apocalypse… job apocalypse being delayed, a Federal Reserve Bank small highlight… highlighted few CEOs need to reduce their workforce, a study by the Federal Reserve Bank of Philadelphia.”

The calibration: “I know a lot of people are feeling job insecurity at all levels of [seniority]. I think we need to do something to address that, and I know many people have been laid off because of over hiring for the pandemic or hiring from zero-interest-rates period. So I’m not saying the job situation is perfect — but the job apocalypse is plausible to talk in some circles, I just don’t see it happening.”

The conclusion: “Which is why I’m eager to keep on investing in all of us becoming better at building these things.”

This is the practitioner-leader-vantage twin of Stanford GSB’s academic-economist-vantage on the same question — Jones provides the theoretical mechanism (weak-links / jobs-as-bundles-of-tasks); Ng provides the empirical headcount observation from inside hiring.

6. Parallel skill development as the framing for both product announcements (~11:07–12:11)

The bridge between the analysis and the announcements. Ng’s framing: “As our coding agents become more capable — and this includes agent skills like the Anthropic agent skills and so forth — our people need the complimentary skills to help us to drive the coding agents in the appropriate way so that we can partner together to build new things.”

Two tracks needed in parallel:

1. Coding agents getting more capable — Anthropic agent skills and the general capability uplift of the underlying models.
2. Human skill development — the complementary skill set humans need to drive the agents effectively.

This frames the two announcements as addressing the two tracks separately: Context Hub for the agent track, Code Dream / Code Realm for the human track.

7. Context Hub (~12:12–14:16) — for AI agents

The agent-track tool. Built by Ng with Vivek Prasad and Sanyam Hota. “It turns out building blocks are proliferating so fast that most of our coding agents, which have a knowledge cutoff date, will often hallucinate or use deprecated APIs or not know about the latest tools you want to use.”

The canonical worked example Ng demos on stage: “Even today, if you ask Claude Code to call the OpenAI API, it will use the chat completions API, which is older and deprecated, and call older models because most of the training data on the internet uses this older API, even though the newer APIs have been out for a long time.”

What Context Hub does: provides up-to-date documentation as context for the coding agent. With Context Hub installed, Claude Code generates code using the newer responses API and knows about GPT-5.5.

The CLI interface demo:

• chob search OpenAI — returns a list of relevant packages.
• chob get OpenAI/chat — fetches the latest documentation for that package (Ng describes the OpenAI/chat docs returning as a 600-line markdown file). “Built for your coding agent to use” — though humans can read it too.

The failure mode it addresses is knowledge-cutoff-leads-to-deprecated-API-calls — distinct from but related to the slop-squatting attack vector SEI surfaced one day later. Both are AI-coding-supply-chain concerns: slop-squatting addresses fake-package-names that don’t exist; Context Hub addresses real-package-APIs that are deprecated.

8. Code Dream / Code Realm (~14:18–17:50) — for human skill development

The human-track tool. “You’ll be the first people in the world to hear about this” (announced ~0:18 of the talk; reveal at ~14:18). “Not a course — a conversation.”

The product surface (described while Ng tried to live-demo over conference Wi-Fi):

• A video-call interface where the user can talk to Ng (or an AI version of Ng) in real time. The AI-Ng has “information you need to try to answer your questions as I will.” User can interrupt at any time via a microphone button.
• A browser-based terminal alongside the conversation. The user pastes commands from the slide (e.g., npm install chob, start up open code, write Python) and practises with modern AI coding agents in the browser without local setup.
• A “Codoji build environment” (ASR-uncertain name) for the hands-on practice surface.

The pedagogical claim: “In lieu of taking a course, you can come on to a video conversation and have a conversation with me or an AI version of me to try to gain new building blocks and code in this more modern AI coding way.”

Available in preview as of the talk date. “The first experience will show you how to use Code Assist app and give us feedback.”

9. The wrap-up framing (~18:02–19:10)

Ng’s two-track recap:

• Coding agents are working well — but give them access to latest building blocks via Context Hub.
• Human skill development — Code Realm conversations as a first-experience in the parallel-skill-development model.

Closing thesis: “The job apocalypse is not being nearly as bad as people are hyping it up to be, and I think it’s important that we keep on learning these building blocks and learning these skills.”

Linked entities and concepts

Entities directly named or substantively discussed:

• Andrew Ng — promoted from Dangling to entity page (this is his second substantive source mention and his first solo-headlining ingest; first mention was the brief “unbig in AI” cross-reference from MIT March 2026).
• DeepLearningAI — channel-entity Dangling first-mention; promote on second source under this author: value (current convention; cf. YC Root Access promotion path).
• Vivek Prasad — Dangling first-mention (Context Hub co-builder).
• Sanyam Hota — Dangling first-mention (Context Hub co-builder).
• Context Hub — Dangling first-mention (Ng’s new tool; concept-page candidate on second-source mention).
• Code Dream / Code Realm — Dangling first-mention (Ng’s new product; the canonical product name is ASR-uncertain — both forms appear in the transcript).
• Anthropic — substantive cross-reference (“agent skills like the Anthropic agent skills and so forth” — Ng frames agent skills as Anthropic-anchored). Source-count: increment.
• OpenAI — substantive cross-reference (the chat completions API vs responses API worked example for Context Hub). Source-count: increment.
• Gemini / Codex / Claude Code / Open Code — referenced as candidate coding-agent tools.
• Federal Reserve Bank of Philadelphia — Dangling first-mention (study Ng cites on the job-apocalypse-being-delayed narrative).

Concepts touched substantively:

• agentic-engineering — Ng’s PM-bottleneck thesis + the cascading-bottlenecks observation + the engineer-plus-PM-collapse-into-one structural endpoint + the 100%-AI-coding norm at frontier teams. Source-count: +1.
• agent-harness — Context Hub as a harness-adjacent tool (provides up-to-date context as part of the agent’s working surface). Source-count: +1.
• vibe-coding — Ng’s 100%-AI-coding personal practice is a vibe-coding-at-frontier-altitude worked example, with the calibrating caveat (“this is not a religion… if you’re working for NASA…”). Source-count: +1.
• ai-employment-effects — Ng’s job-apocalypse counter-claim + the massive unmet demand for AI-engineer-with-generalist-skills observation. Source-count: +1.
• durable-skills — the AI-engineer hiring rubric (coding-agent fluency + building-blocks knowledge + generalist skills) as the operational specification of the 2026 durable-skill mix. Source-count: +1.
• enterprise-ai-adoption — Ng’s AI-native team of generalists prescription + the cascading-bottlenecks observation (design / legal / marketing / sales as the second-wave bottlenecks) is the team-shape-implication of enterprise AI adoption at frontier-team altitude. Source-count: +1.
• automation-vs-augmentation — the PM-becomes-builder + designer-becomes-builder collapse is the augmentation-at-role-level dynamic Ng catalogues. Source-count: +1.

Dangling (single-source mentions, deferred): Andrew Ng (now promoted), DeepLearningAI, Vivek Prasad, Sanyam Hota, Context Hub, Code Dream, Code Realm, Federal Reserve Bank of Philadelphia, Open Code.

Caveats

• 19-minute conference keynote — substantively dense but tightly scoped; many claims (100%-AI-coding norm, PM:engineer ratios collapsing, we can’t find enough of these people) are Ng’s empirical observations from inside his own teams + cohort observations, not externally peer-reviewed.
• The 100%-AI-coding claim is the most aggressive headline-multiplier framing in the wiki’s 2026 corpus; the Wu / Momentic conditional-on-baseline counter-thesis applies. Both anchors should be carried.
• Context Hub and Code Dream / Code Realm are vendor-product announcements with overt promotional motive; the product surfaces described are not yet user-validated by independent reviewers at ingest time.
• The Code Dream / Code Realm product name is ASR-uncertain — Ng uses both “Code Dream” and “Code Realm” on stage; the “Codoji build environment” may be ASR mangling of a product name not transcribed cleanly.
• The Federal Reserve Bank of Philadelphia study Ng cites is named but not URL-anchored; treat as a referenced study, not a confirmed citation.
• The 600-line OpenAI/chat markdown file Ng demos is descriptive (Ng’s count, on stage) — actual file size and structure not independently verified.
• Live demo over conference Wi-Fi (Ng’s own framing: “the Wi-Fi here is not great, so I’m going to see if the live demo works”); audience may have seen visual elements not captured in the transcript.
• The talk’s product-announcement frame means the future-of-software-engineering analysis is at risk of being calibrated to flatter the announcements. Cross-check Ng’s PM-bottleneck observation against Hu / Tan and Schoening for triangulation (filed as supports edges).