1-minute summary
Day 2 traded hype cycles for hard edges: photons instead of electrons for model bandwidth, agents that handle work allocation (not just knowledge), embodied systems learning across fleets, plus sober playbooks for hallucination control, assistive robotics at home, and AI product-market fit. If Day 1 was about the narrative, Day 2 was the operator’s manual – how to compute, ship, and scale when AI touches atoms.
The deep currents
From opening keynotes to late-afternoon firechats, a through-line emerged on Day 2: AI is getting physical, operational, and accountable. It’s not just writing text or code, it’s driving in ports, walking in factories, standing people up, and re-wiring how work gets done.
1. Beyond silicon: photonics as the next AI compute wave
Chip veterans and photonics founders sketched a near-term path where light does the heavy lifting for AI: orders-of-magnitude higher interconnect bandwidth, lower latency between memory and compute, and energy per operation that finally bends the curve. Hybrid designs, CMOS control + photonic data paths, look like the pragmatic bridge from today’s accelerators to tomorrow’s inference fabrics. The honest cautions: packaging, yield, software tooling, and foundry capacity. Expect photonic co-processors to appear first in inference + networking roles, then creep upstream toward training as compiler stacks mature.
“We don’t have to win every time to win the system. If light carries the bandwidth, the whole stack breathes.”
2. Can machines reconnect a fractured web?
Speakers wrestled with the paradox: systems that personalize also partition. The corrective pattern emerging: machine-mediated serendipity, recommendation agents that deliberately inject counter-views, common references, and shared context into feeds, while exposing source trails users can check. The social contract shifts from “watch more” to “understand more”. Expect explainability UX (why you saw this, what’s missing, what others saw) to become a default toggle, not a lab demo.
“Reconnection isn’t ‘more content’, it’s shared context. Machines can supply it, but humans must decide the goal.”
3. Hallucinations: from hand-wringing to engineering
Legal, health, and enterprise builders converged on a working recipe for truthful AI in production:
- Constrain the problem (narrow tasks, strong schemas).
- Ground with authoritative corpora and tool use (search/db/calculator) by default.
- Ensembles and self-checks (self-consistency, verifier models, retrieval-augmented critics).
- Citations by design with abstention pathways: “I don’t know, here’s what’s missing”.
- Human in the loop where it counts, with fine-grained audit logs for procurement and regulators.
“Truth isn’t a temperature setting, it’s a system. Panels of models, verified data, and a graceful I don’t know.”
Tip: Track answerable-rate and evidence-coverage as first-class KPIs, not just accuracy.
4. Embodied intelligence: the fleet is the model
Robotics leaders argued the next “GPT-moment” will look like fleets that learn together. Two building blocks stood out:
- Foundation models for the field: perception + control stacks that generalize across bodies (humanoids, wheeled, legged) and terrains, no GPS, no pre-mapped trajectories, minimal assumptions about connectivity. “One error in 10,000 can be catastrophic”, so the frontier is risk-aware policies that fuse physics with learning.
- Real-world data at home: assistive exoskeletons now pack 12 motors, 27 sensors, ~1.56M datapoints/sec; after 100+ hospital deployments, the push is at-home rehab, where daily use creates the richest gait dataset, and the tightest safety loop.
“Add robot #1000 and you don’t add one worker, you add 999 learning links.”
Enterprise angle: Start where structure + stakes + repetition align (data centers, refineries, yards). Budget for fleet ops + simulation + safety evals as core line items, not afterthoughts.
5. The AI product revolution: opinionated > omnipotent
Investors and founders aligned on what’s actually working:
- Opinionated products that solve the same job thousands of times (support QA, claims, KYC, legal ops).
- Data moats from usage telemetry, domain schemas, and human feedback—not just parameter count.
- Prototype-to-proof in weeks: ship a working slice and measure retained usage before you scale.
- Designers/PMs who code: cross-functional builders collapse cycles and own outcomes.
“Infra is crowded. Workflows with pain + budget + data win first.”
6. After knowledge work: welcome to the allocation economy
Dan Shipper argued that when models write, summarize, and draft, the scarcest skill becomes allocating intelligence, decomposing problems, assigning them to agents, and verifying outcomes. His company now runs five businesses with ~15 people because everyone is “AI-native”, and the CEO still commits code thanks to agentic tooling. The rollout playbook:
- Leaders must use AI daily (no memo leadership).
- Give the top 10% early adopters space to find wins.
- Package those wins (prompts, guardrails, SOPs) for the next 80%.
- Expect flatter orgs as information routing compresses middle layers.
“Stop managing knowledge. Start managing agents.”
7. Assistive robotics: design for dignity, reimburse for reality
The assistive track emphasized human factors + economics: comfort and trust determine adherence; adherence determines outcomes; outcomes determine reimbursement, which then funds better hardware and data. Standards (e.g., ISO 13482-style safety thinking), fall-recovery scenarios, and clear hand-off cues between machine and caregiver are moving from slideware to checklists. The moonshot is mundane and noble: stand, walk, cook, shower – safely.
Spotlight: what the Day-2 binder adds
- Program coherence: Threads across stages – safety, sovereignty, and speed – are not siloed. The same trust primitives (provenance, audit logs, abstention) recur in ads, news, robotics, and finance.
- Operator mindset: “Less deck, more demo” isn’t a slogan, it’s showing up in how teams budget – prototypes, eval harnesses, and post-deployment telemetry get the money first.
What to do next – actionable, not aspirational
- Adopt a hallucination SRE: create an incident playbook (detect → route to human → patch data → regression test) and track answerable-rate + evidence-coverage.
- Pilot an embodied proof: pick one repetitive, structured task (yard shuttling, valve inspection, hot-aisle patrol). Budget for fleet ops + safety eval from day zero.
- Resegment AI bets: kill the omni-bot. Fund opinionated products where you own data + distribution; require a demo with retained users in 4–6 weeks.
- Make allocation a skill: train a pod on agent orchestration (task graphs, verification, rollback). Promote for throughput per head, not slide count.
- Prep for photonic interconnects: ask vendors for roadmaps; earmark a small inference tier to explore photon-assisted bandwidth where caches dominate cost.
- Design assistive pilots for adherence: comfort, fall scenarios, caregiver hand-offs; measure adherence → outcomes → reimbursement.
Why it matters
Day 1 redrew the map – power shifting to the Global South and CEE, trust becoming the differentiator, sovereignty a feature, and “demo over memo” the new tempo – while Day 2 showed how to build for that world. Photonic interconnects, embodied fleets, and agent-run workflows turn narrative into throughput; hallucination control, provenance, and auditability harden products for regulated buyers; and the allocation economy rewrites org design so small, AI-native teams out-ship bigger incumbents.
Put simply: the markets highlighted on Day 1 will reward the operators from Day 2, those who can turn trust into infrastructure, speed into evidence, and sovereignty into a sales advantage.
Catch up with our complete Web Summit 2025 series:
Web Summit 2025 Opens with a Power Shift: Pix, Poland and Sovereign AI Take Center Stage
Web Summit 2025 Day 1: On the Ground in Lisbon – AI Grows Up and Gets Practical
Web Summit 2025 Day 3: Health OS, Query-Based Media and the Race to Zero-Delay Money
