Video summary
Mitä tekoälyltä voidaan odottaa seuraavaksi? | inderesPodi 253
Main summary
Key takeaways
Episode Overview (inderesPodi 253)
The episode revisits and updates earlier predictions about AI. While specific timelines and details can lag, the overall direction—agentic workflows and the rapid automation of knowledge work—has largely played out.
What has (and hasn’t) changed since last year’s AI predictions
- The discussion compares older AI podcasts (including pre–ChatGPT era views from major AI figures) to show that early public forecasts often sounded “naive,” while broad trends still held.
- A key theme is that mainstream attention may make big labs’ leaders more cautious in public statements, even as their organizations continue pushing forward.
“Biggest technological leap” but unclear timelines and messaging
The guests frame AI as an unprecedented technological leap, but they note a wide spectrum of public narratives:
- Some leaders emphasize transformative or even dystopian outcomes.
- Others downplay timelines or claim limited near-term impact.
The host argues that talk about future labor-market disruption is often the most “honest,” while claims about AGI timelines or robot takeovers tend to be more guarded or inconsistent.
AI is not a “technology bubble,” but capital-market expectations may be
They distinguish between two realities:
- Technology reality: capabilities and practical utility are increasing quickly; they believe AI is “vastly underutilized.”
- Market reality: expectations and valuations can still be distorted (i.e., some “bubble” behavior can exist in capital markets).
Conclusion: AI is not merely a bubble. Even if model development stopped where it is, there would still be major remaining value in applying existing tools better and building new workflows.
Agentic workflows: the practical “revolution”
The core operational concept is agents, not just chatbots:
- An agent-based workflow contains language-model-driven decision points that can trigger tool use, retrieve information, and take actions inside systems.
- Examples include:
- retrieving data,
- updating internal systems,
- automating operational tasks (a coffee-mug example illustrates perception → decision → action loops).
The host credits the “agent revolution” for shifting value from simple coding assistance to structured, repeatable loops that can be orchestrated for real business outcomes.
Risk management and hallucinations: why workflow design matters
A practical engineering principle emerges:
- Break tasks into smaller deterministic/verified steps so the system’s uncertainty and hallucinations are less likely to derail outcomes.
- Different use-cases (e.g., creative writing vs. finance/accounting) tolerate different error rates, so tool access and process design must differ accordingly.
They emphasize that “mitigating hallucination risk” is essentially agent development work: designing processes and guardrails around model behavior.
Inference demand is becoming a bottleneck (and changing the chip story)
The episode discusses the idea that inference (compute used when models are running) is growing extremely fast:
- Token consumption is presented as rising sharply year-over-year, leading to concerns about daily capacity exhaustion.
- This shifts the market narrative from “training compute is the whole story” toward “real-time serving demand is the story.”
Efficiency improvements (e.g., using fewer “reasoning tokens”) can extend capacity, but they don’t remove the underlying need for compute—especially as models do more internal “thinking.”
Why training vs inference efficiency still matters (and why hardware competition is intensifying)
They discuss rumors and trends that major hyperscalers build more of their own hardware:
- Nvidia’s dominance is acknowledged, but they expect erosion as companies like Google push TPUs and optimized pipelines.
- They debate Nvidia’s “full data center stack” argument. The host accepts the business logic but notes Google can similarly optimize end-to-end because it controls more parts of the chain.
Cloud/codegen: Claude Code and the “critical mass” effect
They explain why Claude Code / agentic coding tools became popular:
- Earlier agentic coding existed, but Claude Code’s breakthrough is framed as cleaner architecture (less bloat) plus better integration into developer workflows (e.g., command line).
- Adoption was also driven by usability and perceived reliability:
- developers could tolerate or manage mistakes,
- iterative improvements reduced human friction.
They reject the idea that popularity was only due to a specific model release, instead emphasizing product integration, UX, and reaching a “critical mass” where it becomes mainstream.
What’s next in the next ~12 months: persistent agents
They forecast the next step as persistent / learning agents:
- Agents that operate proactively in the background.
- More role-specific agents in business contexts (e.g., sales researcher, finance analyst) rather than one generic “do anything” bot.
They also argue that “genericity” matters less than persistence and bounded autonomy that can learn and execute safely.
Europe/EU angle: why they expect Europe to do well
The host argues the EU can win (or at least compete strongly) in the AI era if it leverages:
- Better process discipline (“by the book” foundations) that agents can plug into well (tests, documentation, clean architecture).
- Data protection, democracy, and human-oriented values as competitive strengths.
- Ongoing capability development (not merely model training volume).
They advocate against “Silicon Valley worship” and against fatalism that the US/China will inevitably dominate.
Listening recommendations and community role
They recommend:
- Lex Fridman podcasts
- Hard Fork
- Darknet Diaries (AI-focused episodes)
- An “Agentics Finland” WhatsApp group as a valuable filter for experiments and high-signal sharing
Presenters / Contributors
- Host (inderesPodi) (name not clearly provided in the subtitles; multiple “Tommi/you” references appear but not definitively identified)
- Markus Haavi — AI and Automation Director at Hoxhunt; AI researcher (AI laboratory)