Don Kim

Tech Lead at crypto exchange
prev. Tech Lead @Alipay, responsible for 160M+ MAU

[Don Kim @Dondyu]

One thing to acknowledge : SG property is generally less volatile compared to S&P. So the risk of wrong market timing is less with SG property. If you invest before a market crash (in either S&P or property) the average long term returns would favour property. Likewise , we can make an argument for timing a bull market, which would favour S&P

[Don Kim @Dondyu]

I am now one of these people 😂 it does get inconvenient walking around with the laptop open tho

[cedric @cedric_chee]

My field report on how frontier organizations are operationalizing AI-native systems @aiDotEngineer SG Day 1

https://t.co/JbmW9GVUQ8

[Don Kim @Dondyu]

@rachpradhan @perplexity_ai Fundamentally, Google has the best search capabilities, and I thought it’s a matter of time till they start leveraging those capabilities in their AI products. The shift is happening, will be interesting how the AI search landscape consolidates over time

[Don Kim @Dondyu]

RAG is fundamentally a search problem, no matter if search is done for agent or for human it’s still the same search. But of course there are some extra optimisations that can be done for agents (AEO) on top of SEO

[Jack Colby @JackColbySEO]

new "SEO is dead" thread every 30 minutes today.

meanwhile google's own dev docs, updated 4 days ago:

"is SEO still relevant for generative AI search? in short, yes!"

"optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

cool, cool.

[Don Kim @Dondyu]

@TejasKumar_ @aiDotEngineer Ngl you cracked me up with your jokes 😄 and also just rewatched your talk on YouTube! 🙏

[Don Kim @Dondyu]

@cedric_chee Thanks! Likewise looking forward to more of your thinking :)

[Don Kim @Dondyu]

Days 2 & 3 at @aiDotEngineer Singapore 🧠. What actually stuck:

1/ LLMs are non-deterministic and therefore unsafe by design.
Evals, classifiers, HITL all reduce risk — they don't eliminate it. Last line of defense: environment isolation. Contain the blast radius. Assume the agent will eventually do something wrong.

2/ Productivity gains from AI have exposed a new set of bottlenecks.
Human collaboration, attention, prompt injection, feature bloat. The constraint is no longer what the model can do. It's everything around it.

3/ Long-horizon tasks fail from goal drift, not token limits. - Z .ai
3 failure modes: goal drift, error accumulation, inability to pivot. Fixes are operational — reread the goal every N steps, verify after each action, ask the agent whether to continue or abandon.

4/ 27.6% of PRs are now AI-generated. — Greptile
AI beats humans on revert rate, but loses on maintainability. Less obvious: PR descriptions are now a prompt injection surface. Code validation needs to scale with agent output volume.

5/ Stripe Minions: 65% of agent PRs for simple code changes approved in one shot.
One dev box per agent. No shared environments.

6/ Good design cannot be trained on outcomes.
Pattern matching ≠ creative process. Because generation is cheap, the new trap is shipping everything you *can* build instead of what you *should*. Craft and restraint are the new differentiators.

7/ Scaling model weights is hitting a ceiling. — Sara Hooker, Adaption
More weights = disproportionate compute cost vs intelligence gains. Era of adaptation: synthetic data + self-improving models + adaptive interfaces. Who adapts fastest wins.

8/ Underexplored: where is enterprise AI adoption actually heading in SEA?
On the ground in Singapore, aggressive enterprise demand wasn't the dominant story. The gap between what's being built and what's being bought felt wider than I expected. Worth watching.

Deeper dives are coming next week.

[Don Kim @Dondyu]

Perfect ending to a perfect conference!

[Wayne Sutton @waynesutton]

Just an inspiring closing @aiDotEngineer Singapore talk from @agrimsingh of https://t.co/bfU3xaxTu4 . @SherryYanJiang and the entire team hats off. Community leading at its finest!

[Don Kim @Dondyu]

Last day at AI Engineer let’s go!!

[Don Kim @Dondyu]

Day 1 at @aiDotEngineer Singapore 🧠

Notes from workshops by Stripe, AWS, Vercel, Convex, Arize

1. AEO is the new SEO — your product needs to show up in AI responses, not just Google.

2. Evals are your moat — without them, switching models = weeks of manual testing. With them = hours.

3. The eval loop that separates production teams from everyone else: Instrument → Trace → Evaluate → Annotate → Analyze → Fix & Deploy → repeat.

4. Agents need money — Stripe is building KYA (Know Your Agent), wallets, and HTTP 402 as a payment primitive. AWS
AgentCore is shipping the production runtime, memory, observability. The agent economy has rails now.

5. Schema-first is non-negotiable. Agents infer the modelling predictably based on schema. Get it wrong early, pay for it forever.

6. DB layer is evolving too — AI agents work well with transactional backends that handle real-time mutations and queries without blocking on 3rd party AI models communicated though websockets.

7. Performance plateaus — after enough prompt/RAG/model experiments, you stop improving on the same axis. The unlock is injecting new ideas to the agents with (e.g. research papers), not more iterations.

The fundamental pieces are in place. The infrastructure is here. The industry just needs to catch up — and keep iterating.

[Don Kim @Dondyu]

Excited for workshops at @aiDotEngineer !!!

[Don Kim @Dondyu]

Loved @GeoffreyHuntley’s talk yesterday. In the age of noise and complexity, few people simplify and talk things fundamentally.

[Don Kim @Dondyu]

@waynesutton @aiDotEngineer @convex @existentializzm @mikeysee SG office let's goooo!

[Don Kim @Dondyu]

@omarsar0 Made an HTML Article Summary using Claude for myself.

Sharing for those short on time!
https://t.co/lXomwE1ff5

[Don Kim @Dondyu]

Thanks to all the speakers and organizers @aiDotEngineer . Great conversations throughout the event.

Picture borrowed from @brianchew 🙏

[Don Kim @Dondyu]

Went to the @aiDotEngineer Unconference event yesterday.

3 small but non-obvious insights stood out to me (thread):

[Don Kim @Dondyu]

3/ We can rewrite open-source libraries for dramatically better performance in lower-level languages.

@rachpradhan built MerlionJS - a NextJS alternative - with 100x+ lower average latency

AI dramatically lowers the cost of rebuilding foundational software.

[Don Kim @Dondyu]

2/ AI summaries of AI summaries degrade fast.

This mirrors my own experience building AI workflows:

One weak upstream AI output can silently derail every downstream node.

The failure compounds.

Worth examining carefully when designing multi-step systems.

@hsukenooi

[Don Kim @Dondyu]

1/ Enterprise sales for AI products are trust-driven to a significant extent.

Anecdotally, these 3 strategies seem to work:

• Open-source the codebase → builds trust
• Founder social capital → credibility compounds
• Benchmark performance → gives enterprises something measurable

Traditional enterprise pitches like: “We improve X metric by Y%” still don’t seem fully realized for agentic systems.

@minu_who