Cross-event comparison · 2024–2026
The same wave, watched from two shores.
What the Agentic AI Summit 2026 said in Berkeley, read against what three AI conferences said in Taipei over the two years before it. All four events have full note sites by the same author — every claim below links back to its source.
Four events, one wave
The lineupThree of these are Taiwan's largest general AI gatherings — the Taiwan AI Academy annual conferences of 2024 and 2025, and the Generative AI Conference in May 2025. The fourth is Berkeley RDI's Agentic AI Summit, the subject of this site. Each card links to that event's full note site.
Aug 1–2, 2026 · UC Berkeley
Agentic AI Summit 2026
Frontier labs, academia and enterprises on agentic AI itself: capabilities, evaluation, security, governance, robotics and the agentic economy.
Sep 9–10, 2025 · Academia Sinica, Taipei
Taiwan AI Academy Conference 2025
Taiwan's industry and research community on scaling AI past the proof-of-concept: agent failure modes, the evaluation crisis, on-premises compute and physical AI.
May 23–24, 2025 · Taipei
Generative AI Conference 2025
Practitioners and developers at the moment generative AI turned agentic: MCP as the new protocol layer, coding with AI, and evaluation surfacing as the bottleneck.
Sep 27–28, 2024 · Academia Sinica, Taipei
Taiwan AI Academy Conference 2024
The "AI for every industry" edition: RAG versus fine-tuning, sovereign AI and Traditional Chinese data scarcity, and agents just beginning to move from concept to deployment.
Side by side
Vitals| Agentic AI Summit 2026 | AI Academy 2025 | GAI Conf 2025 | AI Academy 2024 | |
|---|---|---|---|---|
| Dates | Aug 1–2, 2026 | Sep 9–10, 2025 | May 23–24, 2025 | Sep 27–28, 2024 |
| Where | UC Berkeley campus | Academia Sinica, Taipei | Howard Civil Service International House, Taipei | Academia Sinica, Taipei |
| Host | Berkeley RDI | Taiwan AI Academy | — | Taiwan AI Academy |
| Format | 4 parallel stages, 14 livestreams | 3 parallel rooms | 1 stage, sequential | 5 parallel tracks |
| Volume | 147 sessions · ~5,000 on site + global livestream | ~58 sessions · ~1,500 attendees | 25 sessions · 23 speakers | 58 talks · 71 speakers |
| Who speaks | Frontier labs (OpenAI, Google DeepMind, NVIDIA…), academia, enterprises | Taiwan enterprises, research institutes, government | Practitioners and the developer community | Taiwan industry, government and academia |
| One-line thesis | Capability is no longer the bottleneck — evaluation, environments and governance are | From adopting AI to scaling AI | From answers to actions | AI for every industry — can we deploy it? |
| Notes | this site | taiwan-ai-academy-conf-2025 ↗ | gaiconf2025 ↗ | taiwan-ai-academy-conf-2024 ↗ |
One curve, four readings
TimelineLined up chronologically, the four events read like one story about agents told at four moments — each conference asking the question the previous one left open.
Sep 2024
Taipei
"Should we deploy?"
AI Academy 2024 is still litigating the basics: RAG versus fine-tuning stays unresolved, sovereign AI and the Traditional Chinese data bottleneck dominate (four tokens per million in training corpora), and agents work only where tasks are "clear, decomposable, verifiable" — an X-ray agent climbing from 50–60% to over 90% accuracy by task decomposition. The site's own 2026 retrospective ↗ scores the official predictions: "agents go mainstream" is the one that did not come true.
May 2025
Taipei
"From answers to actions"
The GAI Conference ↗ catches the pivot in real time: the focus moves from chat to agents that act, MCP reaches protocol-layer consensus, and — for the first time in this series — evaluation is named as harder than capability. Taiwan's position is summed up as "solid application layer, thin model layer."
Sep 2025
Taipei
"Why doesn't it stick?"
Three and a half months later, AI Academy 2025 ↗ is already in the trough: agent accuracy collapsing to 33% when too many tools attach, an evaluation crisis (80% of benchmark questions are multiple-choice; real work isn't), compute flowing back on-premises, and small vertical models beating giants on local tasks. The conference's own May-vs-September comparison ↗ tracks exactly how the framing hardened.
Aug 2026
Berkeley
"Measure first, then grant autonomy"
The Agentic AI Summit takes the same diagnosis and builds a program around it: raw capability is no longer the bottleneck, so two days go to evaluation methodology, security, governance-as-code, world models and the agentic economy. The refrain of day two — measure first, then grant autonomy — is the answer to the question Taipei asked in September. See the daily digest.
Where the rooms agree
ConvergenceSeparated by an ocean, a year, and completely different speaker pools, the conferences still reached several identical conclusions — usually without either room hearing the other say it.
Evaluation is the bottleneck — three rooms, one verdict
Berkeley dedicated full sessions on two stages to evals and concluded that benchmark scores are not production reliability. AI Academy 2025 called it an evaluation crisis with numbers: 80% of benchmark questions are multiple-choice while commercial use is open-ended, and only 4 of 11 open-ended evaluations cleared the accuracy bar — "you can't weigh the depth of water on a scale." GAI 2025 had already flagged judging "good enough" as harder than improving capability. Same verdict, three vocabularies.
Agents earned a failure-mode literature
By 2025–26 every room had stopped demoing agents and started dissecting how they fail. Berkeley: Stoica's two fundamental gaps in agentic software engineering, Tworek's failure modes of long-horizon agents, and Credo AI's "earned autonomy" ladder. Taipei: accuracy dropping to 33% with too many tools attached, Cisco counting 8–14 repeat confirmation calls per agentic engine — while back in 2024 agents only worked where tasks were decomposable. The fix converged too: constrain the environment, verify each step, widen autonomy only as it is earned.
Security moved from the model to the system
Dawn Song's Berkeley keynote showed even the evaluation sandbox becoming attack surface; Zaremba reframed safety as fire-style resilience. Taipei tracked the same displacement from a defender's seat: AI Academy 2025 found employee misuse, not model runaway, to be the top corporate risk and worked through the OWASP LLM Top 10, and Ed Chi put it as "safety is A × B × C, not A + B + C" — one weak factor zeroes the product. Notably, Chi is the one speaker who appears in both rooms: safety multiplication in Taipei 2025, personalized universal agents in Berkeley 2026.
Nobody removed the human
No deployment story in any of the four rooms claimed full replacement. Berkeley's day two kept returning to human-in-the-loop and Dan Klein's super-reliability-over-superintelligence; Salesforce argued for employee flourishing. In Taipei, 91App runs "Rule Engine + AI + human," and Delta's generative design gets to "85% roughly right" before a person takes over. The hybrid isn't a transition phase — both shores now treat it as the design target.
Where they don't
DivergenceThe disagreements are mostly not contradictions — they are different questions, driven by where each room sits in the stack and what its economy is short of.
Two ends of the topic spectrum
Berkeley spent real stage time on recursive self-improvement, AGI timelines, world models and payment rails for an agentic economy — topics essentially absent in Taipei. The Taiwan conferences spent theirs on sovereign AI, Traditional Chinese data scarcity (about four tokens per million in training corpora; ~1,300 legally licensed Taiwan images), digital twins and drone supply chains — topics essentially absent in Berkeley. Neither is behind; they are answering different exam papers.
Different engines: capability frontier vs. labour shortage
What pushes adoption differs at the root. In Berkeley the driver is the capability frontier and the race to govern it — how much autonomy can be granted, and how fast. In Taiwan the stated driver is demographic: annual births halved from 400,000 to 200,000, companies stay hundreds of people short even after adopting AI, and a policy panelist put it bluntly — "only two things can save Taiwan: robots and foreigners." Same technology, opposite scarcities: America is short of trust, Taiwan is short of people.
Robots: foundation models vs. bill of materials
Both shores put robots on the main stage and mean different things. Berkeley's robotics thread is about robot foundation models and simulation as the new infrastructure — Fidler's world models going real-time on a consumer GPU within a year. Taipei's is about cost sheets and supply chains: a humanoid costs about NT$4 million to physically test, digital twins matter for trial production, and robotics vendors now ask for URDF files as deliverables. One room trains the brain; the other prices the body.
Compute: pushing the frontier vs. bending the cost curve
Berkeley's infrastructure talks push outward — photonic computing, rack-scale disaggregated serving with shared-memory KV cache, diffusion LLMs redefining token efficiency. Taipei's push inward on economics: on-premises breaks even against cloud tokens at roughly 250 requests a minute over 1.5–2 years, Phison cut a local training bill from NT$30 million to under NT$1 million with NAND flash, and a 24B vertical model beat a 120B general one on Taiwanese law using a sixth of the VRAM. Two ends of the same cost curve — one moves it, the other rides it.
How to read this page
MethodThis comparison is built from four note sites, not four official programs. All four — this one and the three Taiwan sites — were compiled by the same author, so the lens is consistent, but each set of notes has its own inclusion bar: an empty cell means "not in the notes," not necessarily "not said on stage."
The Taiwan sites carry their own comparison pages, which this page borrows method from: AI Academy 2025 compares itself against 2024 and against the May GAI conference, and the 2024 site checks its predictions from 2026 hindsight. For the Berkeley side, every claim links to a full talk note on this site.
Figures quoted (33% tool degradation, 250 requests/minute break-even, NT$4 million per humanoid, four-per-million Traditional Chinese tokens…) are as reported by speakers at each event — see each site's sources page for external verification.