Sebastian Foo

Digital transformation consultant (World Bank, Deloitte, etc). Exploring how AI and other disruptive tech reshape governments, business and work.

[Sebastian Foo @SebastianFooPC]

3/3 This changes what govt leaders should demand from digital resilience plans. Not just system uptime or recovery targets, but tested arrangements for maintaining essential public functions when systems must be isolated.

[Sebastian Foo @SebastianFooPC]

1/3 The critical PHP vulnerabilities exposed a difficult trade-off for digital government. Taking vulnerable systems offline may protect citizens. But it can also interrupt essential services. Cybersecurity decisions are therefore also public-service continuity decisions.🧵

[Sebastian Foo @SebastianFooPC]

2/3 Govts can make the right cybersecurity decision and still leave citizens unable to access a critical service. It could reveal a continuity issue. The litmus test is whether essential transactions can continue safely through alternative channels.

[Sebastian Foo @SebastianFooPC]

AI time savings don't automatically translate into govt savings. It may increase service capacity, improve quality, or reduce costs. Each requires different evidence. Govts should measure realized outcomes, not simply hours saved, before claiming an AI productivity dividend.

[Sebastian Foo @SebastianFooPC]

OpenAI, PerturbAI, and Scripps analyzed mouse-brain data in 32 minutes using thousands of AI agents. The leads still need lab testing. If AI makes hypothesis generation cheaper, deciding which experiments deserve funding could become more valuable than generating more leads.

[Sebastian Foo @SebastianFooPC]

3/3 Define who may use an AI tool, for what purpose, under what supervision, and with what evidence of benefit. Scale successful uses, restrict higher-risk ones, and reassess when material conditions change.

[Sebastian Foo @SebastianFooPC]

1/3 An AI tool can be suitable for teacher lesson planning but unsuitable for unsupervised use by young pupils. Education authorities should assess intended use, age, supervision and foreseeable risks rather than treating approval as a blanket endorsement.🧵

[Sebastian Foo @SebastianFooPC]

2/3 An AI tutor can give safe, accurate answers without improving independent learning. Product safety and educational effectiveness are different questions. Schools need evidence relevant to both, not just engagement figures, user satisfaction or assignment quality.

[Sebastian Foo @SebastianFooPC]

Approving an education AI tool once may not be enough. A model or product update can change its behavior, risks, or suitability for students. Education AI contracts should require change notifications, proportionate reassessment and clear authority to suspend affected uses.

[Sebastian Foo @SebastianFooPC]

Schools may carry responsibility for student-facing AI without sufficient visibility into how it behaves or changes. The procurement test is not just whether safeguards exist, but whether schools can verify them, exercise meaningful controls, and act when problems emerge.

[Sebastian Foo @SebastianFooPC]

4/4 The implication: do not evaluate AI adoption only by output per worker. Also measure whether people can independently debug, explain, and challenge AI-generated work. Productivity and building future expertise are separate management objectives.

[Sebastian Foo @SebastianFooPC]

1/4 The AI jobs debate is focusing heavily on headcount. Firms need a second metric: skill formation. If AI reduces entry-level work, who develops the debugging, judgment, and domain knowledge needed to supervise AI later? 🧵

[Sebastian Foo @SebastianFooPC]

3/4 A separate randomized study of 52 developers learning an unfamiliar Python library found 50% vs 67% on an immediate comprehension test with AI assistance. The small improvement was not statistically significant. It does not establish long-term deskilling.

[Sebastian Foo @SebastianFooPC]

3/3 Low returns need not mean switch off. Redesign may preserve valuable capability; withdrawal may damage an essential service. My test: which feasible option delivers the best outcome for the resources committed from here?

[Sebastian Foo @SebastianFooPC]

1/3 Getting AI into production should not buy it permanent funding. Gartner's survey of 2,002 CIOs and technology executives found only 17% had guidelines for abandoning low-value AI already in production.🧵

[Sebastian Foo @SebastianFooPC]

2/3 At renewal, compare continuing, redesigning, and retiring the system against the best feasible alternative. Count future benefits, operating costs, exit costs, and service disruption. Past spending alone cannot justify the next commitment.

[Sebastian Foo @SebastianFooPC]

3/3 Ownership can still make sense when workloads are large and persistent, utilization is high, or legal/control, latency, or resilience requirements make shared access unsuitable. The principle is sequencing: prove the need before locking in fixed capacity.

[Sebastian Foo @SebastianFooPC]

1/3 The World Bank has announced an initiative to help emerging-market governments access shared, pay-as-you-go GPU infrastructure as an alternative to standalone national data centers. That raises a strategic question: when to rent compute, and when to own it? 🧵

[Sebastian Foo @SebastianFooPC]

2/3 A successful AI pilot can show that a workload is useful in a specific setting. It does not establish durable demand at the full recurring cost of owned capacity. Before building, test funded demand, utilization, and cost per useful workload.