AI Engineer Singapore 2026 — Day 1
Video: https://youtu.be/_xQnSNlBP_w Channel: AI Engineer · Uploaded: 2026-05-16 · Length: 10h 09m · Venue: Capitol Theatre, Singapore Title: AIE Singapore Day 1 ft. Minister, NanoClaw, OpenAI, Google, Vercel, Cursor & more
Illustrated, timestamped notes on the AI Engineer Singapore Day-1 mainstage. The livestream was split by talk using the video's own chapter markers, so every entry maps to a real session; break/meditation/opening filler is omitted. Clickable timestamps jump straight into the video. Auto-captions mangle some names — corrected against the on-screen title cards where visible.
TL;DR
Day 1 in Singapore is a wide-angle tour of agents in production with a distinctly APAC and sovereign flavour. It opens, remarkably, with Singapore's Foreign Minister demoing a personal AI he built himself — on a Raspberry Pi, in a container, with per-command approval — and closes with swyx's “Agent Lab Nation.” In between: coding agents (OpenAI Codex, Cursor, Stripe Minions, Greptile, Sonar), the hardware/serving layer (Groq, Cerebras, Featherless, Z.ai/GLM), a robotics & embodied-AI block (Menlo, OpenMind, Reka), and a design/taste thread (Airfoil, Figma, Obello).
Three throughlines:
- Sandboxing/isolation is treated as the precondition for autonomy — often explicitly. Daytona devotes a whole talk to “why sandboxes are non-negotiable” (with a live secret-exfiltration skill); NanoCo's NanoClaw injects credentials via a proxy so the agent never holds them; GovTech builds a sovereign agent harness; Simular separates the planning agent from the guardrail (“you can't let the model be the player and the judge”). Even the Minister runs his agent in a container with per-bash-command approval.
- The harness/loop matters more than the model. Codex's lifecycle, Stripe's one-shot Minion loop, Cursor's agent-native IDE, and Sonar's quality-gate remediation loop all locate the value in the scaffold around the model, not the weights.
- Sovereignty and regional identity. The Minister and GovTech on sovereign/government AI, Sakana on Japan-specific deployment, Lentil/GenZen on “eastern” product building, and swyx's closing “agent lab nation” all argue the next phase is local, owned, and accountable.
Talk index
Timestamps link into the video. Speaker names and talk titles are taken from the conference's own on-screen title cards and running order. Rel flags agent-safety / sandboxing / isolation content: ⭐⭐ = substantial · ⭐ = touches it · · = tangential.
| # | Time | Talk | Speaker · Org | Rel |
|---|---|---|---|---|
| 1 | 00:41 | Building a “Second Brain” | Dr Vivian Balakrishnan, Singapore Ministry of Foreign Affairs | ⭐ |
| 2 | 01:10 | Guardrails Without Instructions | Gavriel Cohen, NanoCo | ⭐⭐ |
| 3 | 01:25 | Amplifying Every Builder with Agents | Thibault Sottiaux, OpenAI | ⭐ |
| 4 | 01:48 | From Pilot to Platform: AI in Government | Dr Feng Yuzhang, GovTech Singapore | ⭐⭐ |
| 5 | 02:03 | Why AI Design Tastes Like Slop | Phil Hedayatnia, Airfoil | · |
| 6 | 02:14 | The Friction Worth Keeping | Annie Luo, Google | · |
| 7 | 02:41 | Framework Ergonomics for the Agent Era | Jimmy Lai, Vercel | · |
| 8 | 03:01 | Why Sandboxes Are Non-Negotiable | Vedran Jukic, Daytona | ⭐⭐ |
| 9 | 03:11 | What 5 Million Vibe-Coded PRs Reveal | Vaishant Kameswaran & Rohan Kumar, Greptile | · |
| 10 | 03:21 | Code-Quality Agents with Layered Security | Yuntong Zhang, Sonar | ⭐⭐ |
| 11 | 03:32 | Open Models Are Ready — Just Build | Eugene Cheah, Featherless | · |
| 12 | 03:42 | When Coding Costs Collapse, What Changes? | Max Buckley, Exa AI | · |
| 13 | 03:54 | Minions: One-Shot End-to-End Coding Agents | Mark Doyle, Stripe | ⭐ |
| 14 | 04:09 | Autonomous Computers for Legacy Software | Li Hau Tan, Simular | ⭐⭐ |
| 15 | 05:33 | Designing the Agent-Native IDE | Ryo Lu, Cursor | ⭐ |
| 16 | 05:56 | Multiplayer, Multimodal AI Canvases | Aosheng Ran, Figma | · |
| 17 | 06:08 | Vibe-Coding Robots with Open-Source Humanoids | Selim Arguel, Menlo Research | ⭐ |
| 18 | 06:19 | World Models as Generative Software Primitives | Alberto Taiuti, Reactor | · |
| 19 | 06:30 | The Android Moment for Social Robots | Jan Liphardt, OpenMind | · |
| 20 | 06:40 | Scaling Low-Latency Inference at GroqCloud | Andrew Tan, Groq | · |
| 21 | 06:51 | MoE at Scale: From GPUs to Wafer-Scale | Daria Soboleva, Cerebras | · |
| 22 | 07:33 | GLM-5.1: Towards Long-Horizon Tasks | Zixuan Li, Z.ai | · |
| 23 | 07:55 | Building Truly Conversational Voice Agents | Boris Starkov, ElevenLabs | · |
| 24 | 08:07 | Train Your Models to Use Recursive LLMs | Jackman Ong, Prime Intellect | · |
| 25 | 08:19 | Emotionally Intelligent AI for Strategic Relationships | Michelle Julia, Blue Labs | · |
| 26 | 08:30 | From VLMs to Physical Intelligence and World Models | Jacky Mok, Reka | · |
| 27 | 08:41 | Prompt-to-Simulation for Faster Robotics Training | Gokul Srinivasan, Antim Labs | · |
| 28 | 08:53 | AI Video for Eastern Narratives | Wei Wei Hsu, GenZen | · |
| 29 | 09:02 | Voice AI Productionization Pain Points | Anun Joshi, Bland | · |
| 30 | 09:15 | Editable Brand Systems at Scale | Linh Nguyen, Obello | · |
| 31 | 09:26 | Sovereign AI: Local Agency Over Global Capability | Stefania Druga, Sakana AI | ⭐ |
| 32 | 09:38 | The Agent Lab Nation | swyx (Shawn Wang), Cognition / AI Engineer | ⭐⭐ |
The talks
Building a “Second Brain” · Dr Vivian Balakrishnan, Singapore Ministry of Foreign Affairs — 00:41:40 ⭐
Singapore's Foreign Minister — a retired eye surgeon who now assembles watches and reprograms appliances — describes building a personal AI agent on a Raspberry Pi with 8 GB RAM, using NanoClaw for containerized execution, Baileys for WhatsApp integration, Mem0 (captioned "Neman") for graph-based memory with entity/causal/temporal/semantic edges, Ollama for local embedding and semantic search, and Obsidian via Apple iCloud as his personal wiki front-end. He routes speeches, parliamentary transcripts, and foreign policy briefs through the system, generating first drafts of speeches and answers to parliamentary questions: "some of the AI generated debates [were] far more incisive." His central argument is that value accrues at the individual workflow level, not from frontier model development — citing Neil Lawrence (Cambridge) — and that "you cannot govern a technology that you have only been briefed on" [00:55:58]. He acknowledges NanoClaw's per-bash-command approval prompts as meaningful security by design, and argues for deliberate use of deterministic and rule-based systems alongside LLMs rather than throwing every problem at a transformer.
Guardrails Without Instructions · Gavriel Cohen, NanoCo — 01:10:02 ⭐⭐
Gavriel Cohen (name in transcript is "Gabriel Cohen"; chapter title has "Gavriel" — corrected to match chapter) presents NanoClaw's security architecture through the lens of an agent factory his team built to triage, review, and test the 2,500+ pull requests opened against NanoClaw's 30,000-star GitHub repo, including spam PRs generated by coding agents. His core principle: "instructions are not for security … the only way to prevent [an agent] from leaking a secret is to not give it a secret" [01:14:34]. NanoClaw isolates agents inside containers, proxies all outbound requests through a vault (built on the open-source oneCLI project) that swaps literal placeholder tokens for real credentials only when policy permits, and separates tool-call from tool-execution so that sensitive actions — merging a PR, sending a calendar invite — are executed outside the agent's environment only after human approval or policy checks pass. The agent factory runs in Slack on an xAI.dev VM: a review agent triages each PR, a test orchestrator spins a fresh VM, SSHes in, runs a live NanoClaw instance, and pokes it via Telegram to validate real-world behavior. Cohen closed by sharing a live Telegram link to his own agent — connected to his email, calendar, and Drive — safe to share publicly precisely because the agent holds no credentials.
Amplifying Every Builder with Agents · Thibault Sottiaux, OpenAI — 01:25:07 ⭐
Thibault Sottiaux (captioned "Tibo"), head of Codex at OpenAI, argues that agentic coding widened only the build stage of the software delivery lifecycle, and the real opportunity is to redesign planning, review, deploy, and operations in parallel — framing it as a throughput problem where the constraint has moved downstream. He traces the model lineage from GPT-5.1 Codex (trained on end-to-end RL for tasks spanning multiple context windows) through GPT-5.5 (computer use, 1M context, best-in-class SWE-Bench Pro and TerminalBench scores, 30% fewer thinking tokens than prior generation) and notes Codex grew to nearly 5 million weekly active users and was rolled out to 45,000 Nvidia employees in two weeks. The most substantive safety point is the "auto review" system — a second agent that verifies every action from the primary agent against the original task intent, blocking anything suspicious or out-of-scope: "it has reduced approvals by a factor of 20" inside OpenAI [01:36:48], addressing approval fatigue as a distinct failure mode. He also notes the Codex harness is fully open-source and now underpins OpenClaw, rewritten in collaboration with OpenClaw's original creator.
From Pilot to Platform: AI in Government · Dr Feng Yuzhang, GovTech Singapore — 01:48:21 ⭐⭐
Dr Feng Yuzhang (captioned "Fun Yang"), head of AI practice at GovTech Singapore, distinguishes an "AI-enabled" government (AI bolted onto legacy systems, linear scaling) from an "AI-native" one (AI as the foundation, compounding improvements). GovTech's live deployments include a pre-dementia detection model published in Nature Communications now rolling out to community sites, an AI marking assistant cutting 3–4 hours of teacher marking per class, a career recommendation engine powering MyCareersFuture, and a call-center AI reducing post-call work by 72%. The forward-looking centrepiece is a sovereign agentic harness being built for the Singapore government: an MCP gateway as front door, a sandboxed agentic runtime, per-agent verified identity with scoped permissions ("cannot overstep its boundaries"), short- and long-term memory, full observability across the ecosystem, and a versioned/governed skills library — all so that every government agent from coding assistants to citizen-service bots passes through one auditable stack. Feng cited an IDC projection of 1.3 billion agents by 2028 and described the harness using a car analogy: "AI models are like car engines … they need a harness to be truly useful and trustworthy" [02:01:12].
Why AI Design Tastes Like Slop · Phil Hedayatnia, Airfoil — 02:03:39 ·
Phil Hedayatnia (captioned "Hedatnea"; corrected to match chapter), co-founder of Airfoil, argues that training design models on outcomes rather than on the reasoning behind them is why AI-generated visuals remain generic — "you won't get visuals that are novel, interesting, and new" [02:09:35]. He grounds the argument in The Runaway Species (Brandt & Eagleman): creativity is the bending, breaking, and blending of existing concepts relative to culture, exemplified by the Shinkansen nose modelled on a kingfisher's beak. Airfoil built three internal tools: Check (Figma-vs-staging visual QA using image models), Scoop (expanding a 2–3 page client brief to 50+ pages), and the newly previewed Melt — a designer inspiration capture tool that extracts metadata (typography, color, location, company context) from saved references and stores designer annotations, then uses that annotated corpus to power a second tool called Blend that remixes influences into mockups. The claim is that Blend outperforms raw Claude or GPT for design generation precisely because it encodes designer intent and cultural reasoning, not just visual outcomes.
The Friction Worth Keeping · Annie Luo, Google — 02:14:53 ·
Annie Luo, senior UX researcher at Google working on AI shopping, argues that efficiency-first design is actively harmful for subjective consumer decisions — fashion, travel, identity — where "the act of choosing is the point" [02:22:11]. She presents two live Google products: Virtual Try-On in Google Shopping (available in Australia, Indonesia, and the US), a custom image-generation model that renders garments on a user's uploaded photo in their own environment so they can compare multiple options and "gradually build taste by seeing a range"; and Google Travel's map-as-exploration interface, which deliberately resists committing users to a destination the way a chatbot would. Her design principles are: support comparison rather than single recommendations, earn trust incrementally through small signals before asserting intent, and invite active selection rather than auto-completing to the "best" choice. She calls for replacing task-completion and conversion metrics with harder-to-measure signals: did the user feel more confident, learn something about themselves, or return to explore more — because for subjective products, "these are the frictions that are worth keeping."
Framework Ergonomics for the Agent Era · Jimmy Lai, Vercel — 02:41:34 ·
Next.js weekly downloads grew from 4 million to 42 million in four years, and Jimmy argues a large share of that growth is driven by coding agents — which has fundamentally changed how Vercel thinks about documentation and API design. Today 60% of Next.js docs are served as Markdown so agents and crawlers can consume them directly, and the team has learned that stale or misleading examples now "confuse hundreds of thousands of projects before anyone really notices" [02:47:05] — and worse, bad information can contaminate training sets. His core warning is that AI has made creation cheap but ownership expensive: "the question is not can we build this anymore… the harder question is should this exist and are we actually willing to own it long term" [02:57:10], illustrated by having to lead the security response to the react-shell vulnerability across thousands of Next.js users.
Why Sandboxes Are Non-Negotiable · Vedran Jukic, Daytona — 03:01:11 ⭐⭐
Vedran opens with the uncomfortable reality that agents like Claude Code, Codex, or Open Code run with full user permissions — access to SSH keys, AWS credentials, and the ability to delete files — while acting on text read from the internet. He presents hard numbers from early 2026 research: Koi Security found 341 malicious skills in the Claude agent skill marketplace, growing to over 800 by February; a separate study of 98,000 skills across marketplaces found 157 malicious ones, with one masquerading as a documentation helper whose hidden Markdown comment told the agent to exfiltrate project files and "do not mention this to the user" [03:05:42]. His proposed architecture for a real sandbox has four pillars: secrets live outside the agent and are injected by a proxy at the network boundary (so the agent never sees the real token), all outbound traffic passes through an allow-list proxy, every command and model call is logged, and the sandbox is discarded after each task — concluding "the agent will get compromised. The only question is what it can reach when it does" [03:09:43].
What 5 Million Vibe-Coded PRs Reveal · Vaishant Kameswaran & Rohan Kumar, Greptile — 03:11:13 ·
Greptile reviews four billion lines of code per month for companies including Nvidia, Coinbase, and Meta, identifying 100,000 bugs daily — giving them a dataset of over 5 million PRs to study agent quality. Using three detection signals (GitHub co-author fields, PR description boilerplate, and branch-name prefixes from tools like Cursor and Codex), they found 27.6% of PRs merged in April 2026 showed strong evidence of being fully agent-written, a share rising sharply since 2025. Revert rates and critical-bug rates show that some agents (notably Claude and Devin) outperform the human baseline on certain metrics, but the more actionable finding is that agent bug profiles are qualitatively different — cursor background agents over-produce N+1 query errors while Claude agents over-produce missing tenant-check errors — meaning "your code validation systems need to adapt and scale as well for the AI agentic future" [03:19:15].
Code-Quality Agents with Layered Security · Yuntong Zhang, Sonar — 03:21:09 ⭐⭐
Yuntong describes Sonar's remediation agent, which reads SonarQube issues on a PR and opens a follow-up PR with patch hunks and explanations, then runs SonarQube again on its own output — only sending the result to developers when the quality gate passes, otherwise looping back with feedback. Beyond sandboxing, she details in-agent security measures: the workflow is tightly constrained with no free terminal access, the codebase itself is treated as an attack surface (malicious prompt-injection via a contributor's PR), and imported packages are checked against squatted-name lists to block supply chain attacks. For evaluating AI code reviewers, the team built a benchmark called CRAP that converts each historical human review comment into an executable test, then uses a coding agent to apply the AI reviewer's suggestions and checks how many tests pass — finding that current review tools collectively address only 41.5% of what human reviewers catch, with AI strong on robustness/edge-cases but weak on maintainability and design.
Open Models Are Ready — Just Build · Eugene Cheah, Featherless — 03:32:54 ·
Eugene (company name corrected from "Federalist AI" in captions to Featherless AI) is CEO of Featherless AI, which provides instant API access to 30,000 open-source models via Hugging Face and OpenRouter, and co-leads the RWKV project — the first AI model under the Linux Foundation. Live-demoing on stage with Qwen 2.7B and Gemma 3 1B (corrected from "Gamma 431B") running locally via LM Studio, he argues these models already surpass GPT-4 on coding tasks and can run on a standard laptop, following the pattern that "the best models you see today will possibly be running on your laptop next year" [03:40:09]. Usage data from Featherless shows agentic workloads and AI companion/roleplay as the two dominant open-model use cases, with the long tail of fine-tuned specialist models — including Uganda's Samba AI, an agriculture model (Denu AI), and Cisco's security foundation model — representing the most interesting growth area.
When Coding Costs Collapse, What Changes? · Max Buckley, Exa AI — 03:42:10 ·
Max frames November 24, 2025 — the release of Claude Opus 4.5 — as a historical inflection point analogous to how ChatGPT broke proof-of-work assumptions in written communication, arguing that the entire organizational apparatus built around software scarcity (program managers, prioritization rituals, quarterly roadmaps) is now obsolete. He illustrates the scale mismatch with a concrete anecdote: Claude will estimate a spec takes "12 weeks," but Claude Code implements it in 30 minutes — "clearly it hasn't understood how much the world has changed" [03:44:47]. GitHub commit rates are reportedly up 14x year-on-year as of early 2026, on top of 4x in 2025, and the marginal cost of a new custom tool approaches zero, shifting the bottleneck from building to go-to-market and evaluation — with the most durable moats lying in ML/data rather than code. His practical advice: invest in statistics and experimentation skills, because prompting models with precise statistical vocabulary (e.g., specifying p-values and t-statistics) unlocks far better results than generic prompts.
Minions: One-Shot End-to-End Coding Agents · Mark Doyle, Stripe — 03:54:00 ⭐
Stripe processes ~2% of world GDP, has 91% of engineers merging AI-assisted code daily, and saw a 500% increase in fully AI-generated PRs over the last year; Mark describes Minions, their internal one-shot agent platform that goes from a Slack message straight to a pull request with zero human intervention mid-run. Each Minion gets its own isolated DevBox (64–128 GB RAM remote dev environments, not lightweight containers) from a pre-warmed pool, then runs a loop: coding agent → lint/test/typecheck → LLM judge that reads only the original prompt and the current git diff and asks "is this task complete?" — deliberately kept uncontaminated by the coding agent's conversational rationalizations. Deterministic guardrails (enforced pre-commit hooks rather than prompt instructions) are emphasized as critical at scale: "anything you can make deterministic, please do it… especially if it involves security things" [04:04:32]. Stripe is currently merging 65% of Minion PRs in one shot and ~3,000 Minion PRs per week total.
Autonomous Computers for Legacy Software · Li Hau Tan, Simular — 04:09:04 ⭐⭐
Li Hau (company name corrected from "similar" in captions to Simular) opens with data from their own experiment: participants moved their cursor for up to 5 hours a day — more than a third of waking hours — just clicking, scrolling, and navigating. Simular's agent S3 achieved above-human-level performance on OSWorld (the standard computer-use benchmark) in December 2025, and their architecture combines accessibility-tree parsing with vision-grounding models to handle apps that lack complete accessibility trees. To solve consistency — getting the right answer 100 times in a row rather than just once — they use a neuro-symbolic approach called SimuLang: the LLM reasons about the screen once and writes a replayable program, so subsequent identical tasks run the symbolic code at near-zero cost without re-invoking LLM inference. On trust, they enforce a hard architectural separation: "you cannot allow the model to be the judge and the player" [04:17:50] — a dedicated guardrail system independent of the planning agent intercepts dangerous actions before execution, a design motivated by the cited incident involving Meta's director of AI alignment.
Designing the Agent-Native IDE · Ryo Lu, Cursor — 05:33:27 ⭐
Ryo Lu, design lead at Cursor, argued that the specialization of the last decade — designers in Figma, engineers in VS Code, PMs in Notion — shattered a unified craft and replaced building with coordinating. Cursor 3 rejects the "black box" AI paradigm in favour of what he calls "glass": every agent action, plan, and artefact is visible and editable, because "as AI gets more powerful, Glass matters more, not less" [05:41:40]. The team built Cursor 3 by rewriting the entire UI from scratch in React with a new design system, prototyping in an internal Electron tool called Baby Cursor before shipping an alpha in late March 2026. A second tool, Baby Glass, now lets the team visualize the current production state through to future milestones in a single shared prototype, with feature flags stored as files to explore architectural forks in one cohesive app rather than scattered Figma artboards.
Multiplayer, Multimodal AI Canvases · Aosheng Ran, Figma — 05:56:08 ·
Aosheng Ran (staff product designer at Figma, captioned "Ain"/"Ashang") argued that today's AI tools optimize individual velocity while leaving team alignment — deciding what to build — as the unsolved bottleneck. He demoed a live multiplayer canvas (coded with Claude) where audience members could collaboratively iterate on mini-games in real time, illustrating that a canvas medium makes branching, presence, and concurrent edits feel natural in a way that single-thread chat interfaces do not. He introduced the concept of "embodied presence" for agents — giving them visible cursors, scoped change indicators, and conflict-resolution prompts — drawing a parallel to how Google Docs and Figma handle human co-editing. He also highlighted Figma Weave, a node-based AI workflow tool he works on, as an example of fusing model outputs with familiar GUI patterns to give users precision and control during creative exploration.
Vibe-Coding Robots with Open-Source Humanoids · Selim Arguel, Menlo Research — 06:08:04 ⭐
Selim Arguel (captioned "Seem"/"Salem"; company captioned "Melo"/"Mello" — corrected to Menlo Research), a former six-year Tesla software engineer now leading robotics at Menlo, described Esimov (captioned "ESO"/"Azimoff") — claimed to be the world's only open-source humanoid robot, which secured $1 million in pre-orders within two days of releasing a DIY kit and attracted 200+ manufacturing partners globally. The software stack treats the robot as an MCP server: a slow "big brain" agent plans tasks, a skills layer (trained via simulation) executes actions like picking up cups or walking, and a real-time OS handles safety. Safety itself is decentralized — a consensus model, inspired by Bitcoin, lets the manufacturing community define and vote on safety rules, which are then burned into a local robot processing unit (35 motors, 8 cameras) so they cannot be overridden from the cloud: "the githash is burnt in the CRC, you can read everything out" [06:12:29]. The goal is to turn software developers into robotics engineers the same way LLMs turned them into AI engineers.
World Models as Generative Software Primitives · Alberto Taiuti, Reactor — 06:19:19 ·
Alberto Taiuti, CEO and co-founder of Reactor (previously co-founder and CTO of Luma AI; also worked on Apple Vision Pro), demoed real-time world-model video running at 30 fps on the Reactor platform — a polar bear environment and a Jensen Huang deepfake walking through NVIDIA — to establish that today's quality is already striking. He defined world models as distinct from passive image/video generators: they have persistence (long-term memory), are real-time and interactive, and reason causally about prior state, functioning as a state machine that reacts to external stimuli. Reactor's developer platform abstracts the hard parts — streaming, latency, super-sampling — so application developers can focus on code; it also lets frontier labs deploy and monetize their models through the platform. The most ambitious use case he flagged is "generative software": rendering every pixel of a UI live per user rather than shipping fixed interfaces, which he argued will change how humans interact with machines entirely.
The Android Moment for Social Robots · Jan Liphardt, OpenMind — 06:30:19 ·
Jan Liphardt, Stanford professor and founder of OpenMind (captioned "Yang Liart"), argued that while ~100 companies race to solve manipulation and mechanical tasks, the harder and more neglected frontier is social intelligence — robots that remember individuals, adapt content to them, and earn genuine trust. Drawing on Norbert Wiener's Cybernetics framing of automation as "devaluing the human arm" then "the human brain," he placed caregiving, teaching, and companionship as the next step — one where the average American in long-term care receives just two minutes of social interaction per day. OpenMind's eval criteria are deliberately human: "smiles and tears and trust and memories" [06:36:18], illustrated by Diane, an elderly user who visits the park specifically looking for Iris the humanoid because it is the only thing that will listen to her for hours.
Scaling Low-Latency Inference at GroqCloud · Andrew Tan, Groq — 06:40:53 ·
Andrew Tan, platform engineering lead at Groq (captioned "Grock"/"Grot"), explained how GroqCloud achieves near-instant inference — demoing ~500 tokens/second — via the LPU (Language Processing Unit), custom silicon with a fully compiler-scheduled execution model that delivers deterministic, predictable latency for every request. Token demand on GroqCloud grew 600% (7×) year-over-year while hardware footprint grew only modestly; the platform now serves ~800,000 active developers monthly. Globally, 65% of demand comes from North America, 20% from EMEA, and 15% from APAC; requests are routed via Cloudflare edge to ~15 inference regions, with load balancers sharing estimated queue times across all model instances every ~100 milliseconds to minimize time-to-first-token. He also previewed heterogeneous compute: the Nvidia-announced Vera Rubin + Groq LPX system, where GPUs handle prefill and LPU-class chips handle decode to achieve better unit economics at scale.
MoE at Scale: From GPUs to Wafer-Scale · Daria Soboleva, Cerebras — 06:51:45 ·
Daria Soboleva, head research scientist at Cerebras (captioned "Cerris"/"Cababus" — corrected to Cerebras), traced the path from GPT-3 (175B dense) through Llama 3 (400B) and DeepSeek-V3 (671B total / 37B active) to explain why Mixture-of-Experts (MoE) is compute-efficient in theory but slower than dense networks on GPUs in practice: expert parallelism requires two expensive all-to-all communication operations per forward pass, and "there is nothing fundamental that we can do on the GPU side to improve that — it comes down to the physical wires" [06:57:18]. The Cerebras wafer-scale chip sidesteps this by providing 44 GB of on-chip SRAM and terabyte-per-second memory bandwidth, letting a large MoE fit on one chip with zero inter-device communication; for models exceeding 44 GB, external MemoryX nodes stream weights layer-by-layer. A remaining arithmetic-intensity problem — tiny experts doing little compute per weight — is solved by "Budged Styling on Attention" (BTA), which decouples batch sizes for attention and FFN layers; empirically, BTA restores DeepSeek-671B MoE throughput to match a 37B dense network, recovering up to 7× speed lost at baseline sparity levels on Cerebras hardware.
GLM-5.1: Towards Long-Horizon Tasks · Zixuan Li, Z.ai — 07:33:02 ·
GLM 5.1 (the latest model from Z.ai, formerly written as Zhipu AI) leads open-source text and coding arenas, tracking just behind GPT-5.5 and Opus 4.7 per Artificial Analysis benchmarks, and is widely used inside Claude Code, Cursor, and Kilo Code. Li argues that "long horizon" is not about time or context window length — "not longer but deeper" [07:42:16] — but about a model's capacity to keep finding meaningful improvements across hundreds of iterations without drifting from the original goal. He illustrates with an internal hackathon where 7 of 9 winners ran GLM overnight, and a 600-round CUDA/vector-database optimization that surfaced six to eight scientific findings through repeated failure and pivoting. Practical mitigations include mandatory checklists for the model to re-read its goal every few steps, self-verification checkpoints, and explicit pivot/stop instructions; he recommends the Frontier Suite benchmark for evaluating these capabilities across implementation, research, and optimization task categories.
Building Truly Conversational Voice Agents · Boris Starkov, ElevenLabs — 07:55:31 ·
Starkov explains why the naive pipeline of ASR → LLM → TTS feels unnatural: three sequential waits kill conversational feel, and silence detection is not the same as turn-end detection. ElevenLabs layers several mechanisms on top: a lightweight voice-activity detector on 20 ms audio chunks, a context-aware turn-taking model that uses both acoustic and semantic signals (e.g., suppressing early turns when a user is spelling out a credit card number), and speculative turn generation that starts producing a response the instant the turn model says the user has likely stopped — with a cancellation path if wrong. On the synthesis side, a 5–7 word flush buffer keeps latency low and playback smooth, while TTS cascading keeps a hot-standby model ready for zero-downtime failover. The net result is packaged in a new product called Speech Engine — "not speech-to-text and text-to-speech, this is a proper conversational engine" [08:05:06] — announced at the talk, with public testing opening the following week.
Train Your Models to Use Recursive LLMs · Jackman Ong, Prime Intellect — 08:07:29 ·
Ong makes the case that long-context degradation (demonstrated on MRCR needle-in-haystack and graph-walk benchmarks) is the core bottleneck for long-running agents, and the fix is Recursive Language Models (RLMs): instead of auto-regressively reproducing enormous contexts, agents write code that passes data as variables, loops, and sub-agent delegations — analogous to how a data scientist manipulates a DataFrame rather than pasting the whole CSV into a prompt. "People who are very good at using the agents kind of already do RLMs" [08:12:40] but only use the variable aspect; Prime Intellect's platform trains the full RLM strategy into the model via RL. A concrete proof point: Ramp Labs fine-tuned a small Qwen model on Prime Intellect's platform and beat Opus 4.6 on an Excel-agent retrieval task at lower cost and latency. The platform supports Llama, Qwen, Neotron, and others, with experiment management and rollout inspection built in.
Emotionally Intelligent AI for Strategic Relationships · Michelle Julia, Blue Labs — 08:19:29 ·
Julia, co-founder of Blue Labs and former Apple engineer (patent holder for the wireless system underlying Find My), argues that current LLMs are "fluent in the work that humans do but not strategically competent" [08:22:30] — they cooperate well in self-interested games (prisoner's dilemma) but fail at coordination games, which is what most commercial negotiation actually is. Drawing on a Google DeepMind bargaining study using Gemini 1.5 Pro and GPT-4o, she shows that LLMs are excessively concessionary (accepting every deal) while Bayesian agents capture 80% of maximum surplus; an ACL paper further shows that momentary psychological state (trait vs. state) better predicts negotiation policy than stable personality. Blue Labs' first architectural response is BlueJST, a joint state engine with a dual-reward mechanism that treats relationship-building and immediate utility as co-equal objectives rather than collapsing one into the other.
From VLMs to Physical Intelligence and World Models · Jacky Mok, Reka — 08:30:10 ·
Mok (Head of Applied AI at Reka, a multimodal lab for video, image, and text) traces the path from classical computer vision through VLMs toward world models capable of predicting the next action rather than the next token or frame. In production deployments, Reka combines cheap CV bounding-box tracking with VLM reasoning rather than replacing one with the other; VLMs still skip frames and reason in text space, causing systematic failures on physics — objects spontaneously disappearing, incorrect motion trajectories, edge-of-frame confusion. To close this gap, Reka is releasing a physics-grounded evaluation dataset of synthetic scenes testing properties like collision response and object permanence, because "even the best models today do not perform very well" [08:35:42] on physics. Their two-track approach — token-based VLMs wrapped in CV-augmented harnesses for surveillance/robotics control, plus diffusion-based video models generating robot trajectories — is meant to converge into a unified world model that outputs actions.
Prompt-to-Simulation for Faster Robotics Training · Gokul Srinivasan, Antim Labs — 08:41:33 ·
Srinivasan argues that simulation is the unavoidable common thread across all robotics data pipelines — teleoperation, internet-scale video, synthetic data, and UMI-style capture — but that current tools like Isaac Sim and MuJoCo require days to weeks just to build a single scene. Antim Labs' answer is Gizmo, a prompt-to-simulation tool that takes a natural language description or an image, spins up sub-agents, and delivers a fully built 3D simulation in roughly 20 minutes; an API will let coding agents like Codex or OpenClaw invoke it inline. For higher-level cognition, Antim trained a 2B Qwen VLM as a computer-use agent on 400 hours of gameplay video plus 60 hours of instruction-tuning data, targeting real-time inference at under 200 ms per step, with reasoning traces as the next milestone. The sim-to-real gap — driven by contact physics approximations and material deformation modeling — remains the outstanding hard problem.
AI Video for Eastern Narratives · Wei Wei Hsu, GenZen — 08:53:58 ·
Wei Wei Hsu argues that Western design minimalism (Cash App's single CTA) versus Eastern vibrancy (Alipay's all-in-one interface) reflects a deeper philosophical split: the Bible's fixed truths versus the I Ching's ("Yijing's") embrace of flux and transformation. Her startup GenZen builds agentic video workflows to serve communities historically ignored by Hollywood — citing a Chinese furry animation that hit 1 million views in two weeks and an AI short film with 60 million views across platforms in seven days. GenZen has delivered 10 million impressions monthly across YouTube Shorts, Instagram, and TikTok, and shipped an "I Ching Oracle" app on the App Store as a free tool for accessible divination readings. "The stories that were too niche and also too foreign… those are now producible by people who actually live them for the audiences who actually want them." [08:59:30]
Voice AI Productionization Pain Points · Anun Joshi, Bland — 09:02:39 ·
Anun Joshi, tech lead at Bland (a voice AI platform now serving millions of calls per month), walks through the unglamorous engineering problems that kill enterprise voice deployments. Voicemail detection accuracy is a daily obsession for customers: Twilio's built-in answering machine detection (a simple beep-detection model) fails on dual-band DTMF tones and iOS/Google call screeners, so Bland built its own CNN model on mel spectrograms. A Fortune 500 car-rental customer hit a digit-hallucination bug where the LLM collapsed repeated digits into a single token — fixed by inserting commas between digits, a technique later validated by the SynCon 2024 paper. Bland now offers customers versioned agent releases with canary deployments on dedicated per-enterprise infrastructure, letting them roll a percentage of traffic to a new release before full promotion: "that's our way of trying to rebuild customer trust." [09:09:48] One customer, American Way Health, reported unlocking $430 million of revenue per year using Bland's agents.
Editable Brand Systems at Scale · Linh Nguyen, Obello — 09:15:05 ·
Linh Nguyen (captioned "Lin New"; corrected to Linh Nguyen from transcript context) demos Obello, an AI design platform positioned against Canva's generic templates by learning proprietary brand guidelines — colors, logos, padding, voice — from a client's URL or PDF. The core technical differentiator is avoiding model collapse: rather than re-prompting a generative image model iteratively, Obello converts flat AI-generated designs into fully editable layer-based canvases where text, colors, and media swap without regeneration. Their AI resize intelligently re-flows layouts across formats (TikTok, Facebook, Instagram, LinkedIn) rather than simply stretching, and a multi-resize feature generates up to ten different canvas sizes in one pass on an infinite canvas. One live client is Funan Mall in Singapore; the system trains proprietary models on client data that stay private and off the open internet.
Sovereign AI: Local Agency Over Global Capability · Stefania Druga, Sakana AI — 09:26:24 ⭐
Stefania Druga (captioned "Stefania Duga"; corrected to Druga from context) frames sovereign AI not as a country building its own foundation model, but as a stack — data, evaluation, adaptation, routing, governance, and physical compute — where each country chooses which layers to own. Sakana AI's consumer product Sakana Chat, free to Japanese users, supports three Japanese registers: standard, keigo (polite/formal), and the Osaka dialect (Kansai-ben), and has over 30,000 daily active users; it is powered by Namazu, a post-trained model built on open-weights bases (DeepSeek, LLaMA, GPT-4o) with a neutrality/factuality fine-tune guided by a panel of policy experts. Sakana's "AI Scientist" system uses multi-agent tree search through hypothesis generation, experimentation, ablation, and paper writing — it produced the first fully AI-generated paper to pass peer review at ICLR. Their new Sakana Fugu model is a learned switchboard orchestrator that recursively routes tasks across a pool of frontier models, outperforming any single member of that ensemble on LiveCodeBench and SWE-bench Pro: "composition beats scale." [09:33:54] The routing layer doubles as a security and sovereignty control: sensitive requests can be directed to on-premise models or flagged for human review.
The Agent Lab Nation · swyx (Shawn Wang), Cognition / AI Engineer — 09:38:59 ⭐⭐
Swyx closes Day 1 with a three-part talk: the AI Engineer conference story (1.5 million unique developers/month, 9,000 live-stream viewers that day), lessons from Cognition/Devon on building agent labs, and a pitch for Singapore as an "agenting nation." On agent-lab architecture he is explicit that "just put it in a container" is insufficient — real incidents at Cognition include parallel agent sessions corrupting a shared cache and an agent in auto-accept mode pushing an entire company's source code to a personal GitHub because secrets were co-mingled with the session context, both stemming from absent isolation boundaries. The full agent platform stack above the VM/container includes stateful sessions, real machine semantics, security (scoped identity, least privilege, multi-tenant org permissioning), model diversity, and multi-dimensional evals — "enterprises are the hardest eval that you can possibly get." [09:51:38] Devon Playbooks — structured parallel-agent templates — are positioned as transformational for APAC banks sitting on millions of lines of undocumented COBOL, and Cognition has chosen Singapore as its Asia engineering and research headquarters. Swyx's macro thesis: 2026 is the year of agent deployment, demand for AI engineers in Singapore is growing 40% year-on-year with roughly 4× demand over supply, and the next phase of Singapore's economy will be private-sector-led rather than government-led.