Конец токенмаксинга, все вдруг вспомнили про ROI

Overview

The speaker argues that the tech industry is starting to “move away” from indiscriminate neural-network automation and to refocus on ROI and measurable business value.

Key points and analysis

• Shift in industry rhetoric after Uber’s COO comments: The speaker says language has changed after Uber publicly questioned the profitability of using neural networks, alleging that the company spent its annual neural-network budget in the first three months and found no compelling business evidence that the spending produced benefits.

• Inefficiency from “uncontrolled” internal use: Many companies, according to the speaker, gave neural networks to employees for experimentation without evaluating outcomes. A concrete example is employees asking models to do tasks (like small code edits) that could have been done more efficiently manually.

• Cost overruns are now forcing evaluation: As the market becomes “inflated,” inference becomes more expensive than the revenue it supports, pushing companies into losses. Even small cost increases allegedly create outsized negative impact—triggering a “domino effect” of scrutiny.

• LinkedIn and “AI neuropsychosis” as a symptom: The speaker criticizes social-media behavior—titles and self-promotion using “AI” regardless of real competence—claiming it reflects hype rather than productivity gains.

• Companies switching to cheaper providers: The speaker claims some firms are quietly moving to Chinese suppliers/cloud services and away from premium frontier models when the price-performance gap is unjustified.

• Backlash/legal risk from fully automated decisions: They cite lawsuits involving businesses like Pizza Hut and Starbucks, arguing that neural networks making decisions without adequate safeguards can cause significant losses.

• Error propagation in sequential decision-making: The speaker notes hallucination/error rates (quoted as ~5–9%) and argues that when outputs become inputs for later steps, overall error probability rises substantially.

• Preference for “decision support” / hybrid systems: Instead of fully autonomous LLM-driven workflows, they advocate traditional systems where preliminary results are produced and then reviewed/confirmed by specialists—reducing errors and labor while maintaining oversight.

• Neural nets can’t “think” and are brittle outside their logic: The speaker emphasizes that models rely on predefined logic/learned patterns and fail when situations change; therefore, minimizing delegated work and keeping human control is safer.

• Two categories of LLM enthusiasts (in the speaker’s view):

  1. People not involved in development who can’t assess quality but like the idea of building without deep competence.
  2. Domain experts chasing PR/hype and attempting to delegate responsibility while keeping their jobs.

• Recommended practical approach: budget/token caps and selective delegation: They suggest limiting how much work is delegated (e.g., a token budget) and evaluating which tasks truly benefit from automation.

• Example used to illustrate “model mediocrity”: The speaker references Michel Hashmakhtu (HashiCorp) saying a neural network optimized a naive rendering algorithm, appearing “optimal,” but that a specialist-designed solution was better—calling this a “fountain of mediocrity.”

• Quality and slower rollout over hype: They cite Zack? / “ZIK also” (unclear wording in subtitles) and stress the idea that product quality and long-term usefulness matter more than rushing to market with massive automation/“speed” claims.

• Recruiting pressure as another sign of change: The speaker ends by noting more recruiters approaching them with job offers, but leaves the cause unclear.

Overall conclusion

The video’s main message is that costs, errors, legal exposure, and lack of demonstrated business value are pushing companies to scale back “token-maxxing” and broad neural-network automation. The speaker advocates measured deployment tied to ROI, hybrid workflows with human oversight, and selective use of AI rather than hype-driven automation.

Presenters / contributors mentioned

• Uber COO (named only as Uber’s COO)
• Sam Altman
• Dario Amodei
• Michel Hashmakhtu / Michel Hashmotu (HashiCorp founder/creator, per the speaker)
• Hashshikorp / Hashjackorp / HashiCorp (company referenced; subtitle errors noted)
• ZIK / “ZIK also” (person referenced, name unclear in subtitles; likely the creator/interviewee of a related discussion)
• iClaw / Anthropic / OpenAI / Claude (organizations/models referenced; not presenters)