🗂️ Field Notes

Real ground-level experience shared by Singapore's AI startup community. Clustered by topic, anonymised, here to help you avoid common pitfalls.

💡 These notes come from real conversations in Singapore's AI startup community, structured and anonymised before publication. Each entry tags the company profile and applicable context. Treat them as reference, not prescription. For policy specifics, defer to the latest official documents.

Research monthly National University of Singapore Agency for Science, Technology and Research Nanyang Technological University EDB EP application Tax compliance Grant programmes Company incorporation

Singapore AI Research Monthly (2026-08)

📅 2026-08 📖 OpenAlex + our curation
Research monthly National University of Singapore Agency for Science, Technology and Research

AI safety & evaluation

  • · [National University of Singapore] Mapping LLM Capability Frontiers via Formalized and Calibrated Probes — NUS-led X-RAY probes LLM reasoning through formally verified, structurally controlled tests rather than raw task accuracy, isolating properties like constraint interaction and reasoning depth. This tackles a core frontier problem in LLM evaluation and interpretability that matters for both capability benchmarking and safety assessment. (https://doi.org/10.1145/3770855.3818029)
  • · [Singapore University of Technology and Design] Auditable Release Control for Pedagogical Leakage in LLM Tutors — SUTD formalizes premature answer disclosure in LLM tutors as "pedagogical leakage" and builds an authorization-aware release control with replayable failure traces, tested against Gemini 3. It's a concrete safety mechanism with inspectable checks, not just a benchmark score. (https://openalex.org/W7172559148)

AI for science & health

  • · [Agency for Science, Technology and Research] A clinical multimodal vision language foundation model with fine-grain explainability — A*STAR built a clinical multimodal vision-language foundation model with fine-grained explainability, pairing frontier multimodal architecture with the interpretability clinicians need to trust AI output. It signals Singapore's push toward home-grown biomedical foundation models rather than adapting generic systems. (https://doi.org/10.5281/zenodo.21772694)
  • · [Nanyang Technological University] The Words of Proteins: Motif-Level Language Modeling for Interpretable Protein Generation — NTU proposes motif-level language modeling for protein generation, using structural motifs instead of raw residues as the modeling unit to make generative protein design more interpretable. It brings LLM-style architecture choices into protein foundation models, an area Singapore is building presence in. (https://doi.org/10.1145/3770855.3819054)

Multimodal & vision

  • · [Agency for Science, Technology and Research] Improving Temporal Action Segmentation via Constraint-Aware Decoding — A joint A*STAR-NTU-IHPC team introduces constraint-aware decoding to improve temporal action segmentation, a core video-understanding task underlying robotics and surveillance pipelines. It's a solid, multi-institution SG computer-vision contribution rather than a single-lab incremental tweak. (https://doi.org/10.1007/978-3-032-31452-9_5)
  • · [Agency for Science, Technology and Research] Causal Scaffolding for Physical Reasoning: A Benchmark for Causally-Informed Physical World Understanding in VLMs — A*STAR and NTU built CausalPhys, a 3,000-plus question benchmark with expert-annotated causal graphs testing whether vision-language models reason about physical causality or just pattern-match. It targets a known weakness in current VLMs with an interpretable, causal-graph-grounded evaluation metric. (https://doi.org/10.1145/3770855.3817582)

Systems & efficiency

  • · [Agency for Science, Technology and Research] Standing Peg-in-Hole Insertion: Demonstrations, Trained Policies, Evaluation Episodes, and Simulation Environment — This A*STAR robotics bundle isolates whether perception or policy causes failures in sub-millimeter-clearance peg-in-hole assembly, releasing demonstrations, trained policies, and a simulation environment for reuse. It's a concrete open infrastructure contribution to embodied-AI and robot learning, aligned with Singapore's advanced-manufacturing AI focus. (https://doi.org/10.5281/zenodo.21287317)

LLMs & agents

  • · [Nanyang Technological University] Toward Plasticity-Preserving KL Regularization for Capability Retention in LLM Reinforcement Learning — NTU researchers target a core LLM post-training problem: RL fine-tuning that erodes capabilities the base model already has. Their correctness-conditioned KL constraint is a methods contribution to how frontier models get trained, not a downstream application. (https://openalex.org/W7172557801)

Other

  • · [National University of Singapore] Copyright Law and Generative AI — An NUS-authored comparative law book on how copyright exceptions in the US, China, EU, Japan, UK, and Singapore apply to generative AI training and outputs. It speaks directly to the legal uncertainty regulators and AI developers are working through, which matters for a policy-focused observatory. (https://doi.org/10.4324/9781003655602)
💬 Takeaway: Singapore's August research trajectory shows a maturation from general-purpose model scaling toward domain-ready, interpretable AI systems: the month spans formal LLM evaluation and safety methods (X-RAY probes, plasticity-preserving RL, pedagogical leakage detection), clinical vision-language models with fine-grained explainability, protein foundation models, and open-source robotics infrastructure. This reflects a strategic shift toward building AI systems that clinicians, regulators, and manufacturers can inspect and trust—anchored by multi-institutional collaboration (A*STAR, NUS, NTU, SUTD, IHPC) and policy analysis (copyright frameworks)—rather than chasing frontier capability metrics.

Singapore AI Research Monthly (2026-07)

📅 2026-07 📖 OpenAlex + our curation
Research monthly National University of Singapore Nanyang Technological University

AI safety & evaluation

  • · [National University of Singapore] A Practical Guide to Interpretability Metrics for Chain of Thought Reasoning — A NUS-authored survey systematically organizes the fragmented landscape of chain-of-thought interpretability metrics, directly addressing whether LLM reasoning traces are faithful rather than just plausible-looking. This is core AI safety/evaluation work on a problem central to trusting frontier reasoning models. (https://doi.org/10.5281/zenodo.21127002)
  • · [Nanyang Technological University] Are heterogeneous graph neural networks truly effective for node classification? A causal perspective — This NTU-led paper uses causal inference to test whether heterogeneous graph neural networks' reported gains on node classification are genuine or confounded by dataset artifacts, a rigor-focused contribution published in the respected journal Knowledge-Based Systems. It represents substantive Singapore-led AI methodology work rather than a downstream domain application. (https://doi.org/10.1016/j.knosys.2026.116595)
  • · [Singapore University of Social Sciences] Governance-Aware Agentic AI for Enterprise Engineering Systems: A Design-Science Reference Architecture and Quantitative Risk-Control Model — Authored at the Singapore University of Social Sciences, this paper proposes a design-science reference architecture and quantitative risk-control model for governing agentic AI in enterprise settings, covering auditability, tool-invocation control, and human escalation. As enterprise agentic AI adoption accelerates, the governance framing is directly relevant to Singapore's AI policy ecosystem, though it remains a conceptual preprint (Zenodo) without empirical validation. (https://doi.org/10.5281/zenodo.21264296)

LLMs & agents

  • · [Nanyang Technological University] Beyond Textual Repository Exploration: Dual-Modal Structural Reasoning for Agentic Issue Resolution — NTU researchers propose DUALVIEW, a dual-modal (text + code-structure) reasoning approach for agentic issue resolution that tackles a known failure mode of coding agents navigating large repositories. It's squarely in the fast-moving agentic-coding space that dominates current LLM agent research. (https://openalex.org/W7167379738)
  • · [Singapore Management University] Knowledge-State Generative Agents for Pre Assessment Question Evaluation — SMU researchers built and empirically validated LLM-based generative agents that simulate students with varying mastery levels to evaluate assessment question quality, tested on real data from 424 students. Published in a top Information Systems journal, it shows credible applied LLM-agent research with clear educational-ecosystem relevance. (https://openalex.org/W7164709949)
  • · [National University of Singapore] ChakapBot: A Generative AI-Powered Chatbot for the Revitalisation of Baba Malay — ChakapBot, built at NUS, is a generative AI chatbot trained on a community-curated corpus to support revitalization and documentation of Baba Malay, an endangered Singapore heritage language, validated in a 26-participant pilot. It's a distinctive, socially-grounded LLM application with clear ecosystem and cultural-preservation significance beyond typical commercial use. (https://doi.org/10.3390/languages11070145)

Multimodal & vision

  • · [Nanyang Technological University] LongVQUBench: Benchmarking Long-Term Video Quality Understanding of Vision-Language Models — LongVQUBench is a substantial new benchmark (1,200+ videos, 1,500 QA pairs) from NTU for evaluating long-term video quality understanding in large vision-language models, filling a real gap since prior benchmarks only cover short clips. It directly tests frontier multimodal model capabilities on temporal reasoning. (https://openalex.org/W7167289797)
  • · [National University of Singapore] NoPA: Non-Parametric Online 3D Scene Graph Generation — NUS-led NoPA advances real-time 3D scene graph generation by replacing coarse single-Gaussian object approximations, improving both speed and geometric fidelity for embodied/robotic perception. It's a solid technical contribution to a core multimodal perception problem with practical robotics relevance. (https://openalex.org/W7167290256)

AI for science & health

  • · [Nanyang Technological University] Reaction-aware molecular representation learning: Toward generalizable artificial intelligence for enzymatic catalysis — NTU researchers introduce a reaction-aware molecular representation learning framework aimed at generalizable AI for enzymatic catalysis, published in Acta Pharmaceutica Sinica B. It targets a core AI-for-science bottleneck—transferable representations of chemical reactions—rather than a narrow, incremental use case. (https://doi.org/10.1016/j.apsb.2026.06.033)
  • · [Agency for Science, Technology and Research] CycPeptMPDB-4D: Multi-Solvent Conformational Ensembles for Predicting Cyclic Peptide Permeability — CycPeptMPDB-4D is a large multi-institution (A*STAR, NUS, NTU) resource of molecular-dynamics-derived conformational ensembles for 5,160 cyclic peptides, built explicitly to train 3D/4D deep learning models for membrane permeability prediction. It's a foundational dataset contribution bridging physics-based simulation and AI-driven drug discovery, reflecting Singapore's growing AI-for-science infrastructure role. (https://doi.org/10.5281/zenodo.21237441)
💬 Takeaway: Singapore's July papers reveal a strategic pivot from frontier model development toward agentic AI systems and AI-for-science infrastructure, paired with growing emphasis on governance, interpretability, and rigorous evaluation methods. The portfolio—spanning real-time 3D perception agents, applied educational agents, molecular discovery datasets, and enterprise governance frameworks—positions Singapore distinctly as both a trustworthiness evaluator and an architect of AI deployment infrastructure for high-stakes domains.

Meeting Notes: AI Startup Meets EDB

📅 2026-02 📖 Community contribution
EDB EP application Tax compliance Grant programmes Company incorporation
🏢 Company profile: AI platform company, ARR approaching $100M, incorporated in Singapore, team mostly remote, around 10 people.

Company classification: shareholding structure, not founder nationality Structure, not passport

  • · EDB applies a clear test for "Singapore company" vs "Chinese-funded enterprise": where the entity is incorporated, whether the cap table contains Chinese entities, and whether customers and revenue are globally distributed.
  • · If the company is incorporated in Singapore, has no Chinese entity on the cap table, and runs a business primarily oriented to North America, Europe or globally, it counts as a Singapore company.
  • · This classification determines whether you go down the headquarters-track route or require a special-handling pathway.

EP applications: going through the front door beats relying on connections

  • · EP thresholds for core executives such as CEO/CTO are not high, and do not require prior hiring of local staff.
  • · Salary must be paid out of the Singapore entity.
  • · About 90% of rejections are due to errors in the application, not policy issues.
  • · Headcount of 10–20 is the most permissive band; once you scale, diversity requirements apply.
  • · Recommendation: apply yourself online first; only engage an agent if you get rejected.

The actual red line: monocultural teams Diversity is not window-dressing; it is a substantive requirement

  • · Singapore does not require a fixed local-headcount ratio, but it cares a lot about team diversity.
  • · Unacceptable: an all-China team, an all-India team, an all-US team.
  • · Acceptable: China + North America + Europe + Australia.
  • · Monocultural companies rarely become truly international — this is a long-running observation from experience, not political correctness.

Counterexample: Manus is a special case

  • · EDB raised Manus on its own initiative, but made clear that it was a special case.
  • · Manus faced an external-regulatory time window and had to relocate its core team in a very compressed timeframe.
  • · EDB engaged deeply: coordinating in advance with the Ministry of Manpower and breaking the relocation into batches.
  • · The expectation was that part of the original team would have to be left behind.
  • · Bottom line: ordinary companies should follow the ordinary process, where success rates are higher. Special assistance is reserved for situations where there is no alternative.
  • · Follow-up (27 April 2026): China's NDRC blocked Meta's US$2B acquisition of Manus on national-security grounds, drawing three red lines (technology sovereignty, data sovereignty, national security). The "time-window pressure" and the "core-team relocation" EDB referenced were, in hindsight, downstream of source-country regulatory direction — re-domiciling alone is not enough to exit source-country jurisdiction, and the "Singapore washing" play was, for the first time, explicitly rejected.
  • · Follow-up (11 August 2026): Manus wrote to users announcing a return to independent operations, formally unwinding the Meta acquisition; user data generated under Meta's ownership (from 29 December 2025) is being deleted. The case has now run its full course from veto to unwind, with Manus still headquartered in Singapore.

Real-world sequencing: compliance before tax incentives Compliance first, incentives second

  • · For a company approaching $100M ARR, EDB recommends the following sequence:
  • · Step 1: tax compliance — international tax plus transfer pricing, with a clear story for why revenue sits here and why profits are allocated this way; engage one of the Big Four directly.
  • · Step 2: define the headquarters positioning.
  • · Step 3: only then consider tax-rate incentives.
  • · Tax-rate incentives: 17% → 15% (around 15 staff), 17% → 10% (25 staff, achieved by year 5).
  • · Incentive policy rewards structures that already work; it is not designed to rescue broken structures.

Two high-leverage programmes

  • · R&D grants: a good fit for technical teams of up to 10 people, available for genuine R&D work tied to core technology.
  • · Local fresh-graduate training subsidy: hire five local fresh graduates full-time, mentored by the CTO or core engineering staff, with the government subsidising part of the salary.
  • · Enterprise Compute Initiative: the government subsidises local companies' use of AI infrastructure providers; you need a Singapore-based team to participate, and it can be combined with AWS, Microsoft or Google.

Company incorporation and directors

  • · It is common for the Singapore entity to be incorporated with an agency-supplied nominee director while the founder still has no EP.
  • · Workflow: use the Singapore entity to apply for your own EP, then assume the directorship after the EP is granted.
  • · Directors carry legal liability.
  • · You can run the process yourself or delegate it to a third-party agent.
💬 Takeaway: Singapore's policy stack is built for companies that already work — and don't want to crash on structural issues. If you're still validating PMF, the government doesn't matter much. But once your risk shifts from "can we make this work" to "structure, compliance, long-term sustainability", this kind of conversation starts to pay off.

Have field experience to share? Contribute via GitHub.

All submissions are structured and anonymised before publication.