林鉦育：Need for Speed — 人要比 AI 兇

1. In one sentence

林鉦育 used one failed and one successful rollout of AI tooling at E.SUN Bank's Intelligent Banking Division to show that in a heavily regulated industry, whether AI actually lands has never been about whether the tool is fast enough — it is about whether people, managers above all, are willing to push process, organisation and their own bosses harder than the AI pushes them.

2. The session

Opening: what one sponsor backdrop meant

He opened with a photo, saying this was the busiest thing he had worked on in the first half of the year: E.SUN Bank "finally got onto the backdrop". He stressed it was not a composite — they genuinely stood on the kind of sponsor backdrop normally reserved for supply chain vendors. On verification, this refers to E.SUN Bank becoming NVIDIA's first financial-services partner at NVIDIA GTC Taipei 2025 and presenting its AI results publicly at the event; GTC Taipei 2025 fell roughly two days before he went on stage (TechNews report). To him the backdrop was itself the most direct evidence that "financial services really has entered AI seriously" — if an industry outside the supply chain can get put on stage by NVIDIA, then it is actually happening, not marketing language.

What "people have to be fiercer than AI" actually means

The line is not about being fierce toward the AI. It is that AI tools move fast, and people who do not actively keep up get left behind by the tool's own speed. His example was his own: as a technical manager who once ran a team of seventy, human code review could not possibly keep up with the code produced each day; but equally, if every engineer accelerates their development with AI and the manager does not use AI to speed up their own review and judgement, management becomes the bottleneck. So "fierce" is about who holds the initiative on pace: the faster you adopt AI, the more people have to actively drive process redesign, ask management for resources, and design the use cases — rather than passively waiting for the tool to handle it.

Two rollouts: you fail once before you succeed once

This was the "one plus one" metaphor at the centre of the session. E.SUN Bank's Intelligent Banking Division actually ran two rounds of AI tool adoption:

• The first, earlier round: only senior developers were asked to test it, in a small group, because "they don't have much time". The test happened, but a few months later the stocktake concluded "nothing changed" — the tool had been bought and tried, but development behaviour had not actually changed, so the money was wasted.
• The second, from 22 August 2024: the division's fifty-plus developers began a two-month full trial of GitHub Copilot, which has since continued for close to a year. During the trial it produced an average of sixty coding suggestions per developer per day, close to forty thousand in total. Of the users, data scientists and machine learning engineers were 35% each, systems engineers 22%, and data engineers and project managers 4% each.

His caution: the first failure was not the tool's fault. It was the absence of a measurement mechanism and the absence of any real change in how everyone worked. The second round succeeded because it was paired with an automated pipeline and quantitative metrics from the start.

Four use cases

What the division's staff actually used GitHub Copilot for concentrated into four categories: data handling (wrangling DataFrames, checking source formats), code generation (HTML, docstrings, test cases, SQL queries, charting code), code comprehension and explanation (explaining Python and SQL, helping work through logs and error messages), and code optimisation (speeding up SQL queries, improving Python function runtimes). He considers these four to cover most of an engineer's daily work.

Security and regulation: the ceiling on AI in finance

He did not lean on this part, but it is the piece that most needs pinning down. Adopting a cloud-based AI tool like GitHub Copilot in financial services inherently raises whether company code, system architecture and even customer data pass through an external service — the source of what he called the industry's particular challenge. Verification fills in the relevant regulatory context: handling customer and employee personal data falls under the Personal Data Protection Act (now administered by the Personal Data Protection Commission); banking as a whole is supervised by the Financial Supervisory Commission, with separate rules for information security and outsourcing. Whether the FSC has published a dedicated core guideline for "AI use in financial services", and what its official link would be, remain unclear — no reliable official page could be found; unverified. What is certain is that every automated flow he described has matching security monitoring and code quality monitoring built in, which is the substance behind his remark that "you can't just say these tools are wonderful and stop there" — monitoring is a mandatory companion to AI adoption in this industry, not an option.

3. Figures and cases

Qualitative results (user survey; speaker's own account, no external source)

Quantitative results (speaker's own account, no external source)

A SQL case study: the same query logic written with and without an index gives identical results and wildly different performance — the version without the index does a full database scan, taking about 450 milliseconds and scanning a million rows; the version with the index takes about 8 milliseconds and scans only ten thousand, more than fifty times faster with a hundred times fewer rows scanned. He used this to make the point that differences in code quality are sometimes completely invisible to the eye, but an AI-assisted tool can catch them at the moment of writing.

Cost-benefit arithmetic (speaker's own account): he converted the performance gain straight into a percentage of salary — assume an engineer on NT$50,000 a month; a mere 10% efficiency improvement saves NT$5,000 a month, far more than the roughly NT$600 a month a subscription to any of these AI tools costs (ChatGPT Plus, Perplexity, Cursor Pro, GitHub Copilot and so on). On checking, INSIDE's feature converts this into a roughly eight-fold return on investment, and adds two figures he did not emphasise out loud: within one month of the trial, total code output rose 50%, and the development team averaged over a thousand lines of code a day. This is the most direct arithmetic he used to get management to back an AI budget.

Development time comparison (speaker's own account): he demonstrated live, generating a complete small project with a front end, an API and an AI model from a single command in about five minutes. He reckons a traditional developer, unless they simply found a ready-made project on GitHub, would need three to four days to write it from scratch.

4. Lines worth keeping

• "People have to be fiercer than AI."
• "To drive any AI change, we fail once first, and then we succeed."
• "It looks like nothing until you look — and then every single item is 10% or more."
• "Everything is a cost-benefit question."
• "Having the AI tool at all is step one of everything."
• "Let's lie flat for another year, because chasing it constantly is exhausting; every time you wait a little, something new shows up."

5. Tools and terms mentioned

• GitHub Copilot: the AI coding assistant adopted across E.SUN Bank's Intelligent Banking Division, and the core case of this session.
• Agent mode: he mentioned the new phase then rolling out, hinting that later sessions would focus more on agents.
• CI/CD and automated pipelines: infrastructure he considers a prerequisite before adopting AI, because without it you cannot measure the improvement.
• Technical debt / code smell: software engineering measures of maintainability; he used the decline in both as evidence of the rollout's effect.
• Personal Data Protection Act: the primary legal basis for handling customer and employee data in financial services — link above.

6. Wider observations

The most striking thing here is that he never presented AI as a cure-all; he spent half the session on the first failure. That honesty is more persuasive to an audience — if you only hear the success stories, it is easy to assume that buying and installing the tool produces the effect. He said outright that "the tool got bought but nothing changed" is the default, and that measurement and process redesign are the deciding variables. For anyone trying to push AI tooling inside their own company — not just in finance — that is a very practical warning: work out how you will measure it and how you will force the team to change its working habits before arguing about which tool to buy. "People have to be fiercer than AI" sounds like a joke; it is actually about who holds the initiative in organisational change.

7. Sources