newsfilter
← Interviews
AI Engineer23 min

I love you so jevving much — Diogo Almeida, TypeSafe AI

Diogo Almeida, Jev

Conference Presentation
  • AI is predicted to become transformative and "so fucking back," rejecting the "saspocalypse" narrative in favor of a "Sassapalooza" driven by machine-native intelligence that adds a fourth semantic logic gate to software, similar to the blue LED revolution.
  • A shift from assistance-optimized models to automation-focused systems is expected within six months to a year, moving the industry question from "whether we can automate" to "how much work is it" by overcoming the "tyranny of the KV cache" and enabling state sharing and non-naive routing.
  • "Real software engineering" will replace "vibe coding" and "LLM wrappers" through decomposition, verified stages, and state management, while "software research" will emerge as a decentralized movement to explore boundaries akin to the early Internet.
  • Specific technical evolutions include semantic-dependent typing allowing local property declarations in lint checkers, the return of voice interfaces via agents that understand intent without hard-coded rules, and real-time intelligence security becoming standard for corporations once sufficiently fast, cheap, and reliable.
  • Performance gains are attributed to "doing the right task" rather than just scaling, with "Jev-like" reliability models providing consistent intelligence to mitigate the "occasional smarts" and "unbelievable" inconsistency of current Large Language Models.
  • The industry faces risks from "slop" generated by externalized costs and bad actors, which can be mitigated by filtering, focusing on quality, and ensuring AI reliability for decisions that incur actual costs.
  • Data is asserted to matter more than compute, and the "bitterest lesson" in the field is optimizing for the correct task rather than relying on algorithmic scaling, a shift that previously enabled GPT-3's success over masked language modeling.
  • Benchmarks are criticized as unhelpful hurdles that hurt those unable to focus on them, while "cursed software" and playful hacker ideas are encouraged because 1% of such endeavors can lead to super awesome outcomes.
  • The current reliance on RLHF is noted as a limitation requiring human verification, with 100% of current models optimized for assistance rather than the deeper capabilities required for true automation.
  • The field must overcome "LLM Stockholm syndrome" and the cognitive dissonance of high benchmark scores versus poor real-world automation to achieve the "reliable intelligence" necessary for a fourth logic age.