Jeff Dean: Trends in Machine Learning [video]
Claims from Jeff Dean’s talk on “exciting trends” in machine learning prompt debate over how much of modern AI is really new theory versus rebranded convolution, with many arguing that transformer-based LLMs and scaling laws dominate practice while fundamental understanding lags. Commenters are split between enthusiasm for large multimodal models like Google’s Gemini and skepticism that current, compute-hungry architectures have plateaued or are overly shaped by corporate incentives and advertising use cases. A major practical theme is the rise of AI tools for summarizing long videos and texts, which some see as a genuine productivity boost and others view as shallow, error‑prone shortcuts that risk flattening depth and nuance in how people learn.
Debate: “All machine learning is convolution”
- Some argue any matrix multiplication (including in Transformers and MLPs) can be seen as a convolution or integral transform, so “everything is convolution.”
- Others push back: only explicit convolution layers count; dot products in high‑dimensional spaces are not typically called convolutions, and nonlinearities break strict equivalence.
- Linked work claims convolutional architectures are universal, but commenters note that “can express” ≠ “is.”
DermAssist and medical AI tools
- Talk mentions Google’s DermAssist as a live product, but users encounter waitlists and 404s; availability in the US is explicitly limited.
- Commenters see repeated overpromising and underdelivery from dermatology AIs in real clinical settings and urge skepticism.
- Other medical AI examples are shared as more successful, alongside an open‑source derm tool.
AI-based video summarization workflows
- Many use LLMs to summarize long talks before watching, citing time savings and triaging value.
- Others prefer video descriptions, skimming, or reading books directly, arguing AI summaries feel superficial and homogenized.
- Multiple tools and personal projects are shared (YouTube-centric summarizers, timestamped topic maps, flashcard generators).
- Some report that LLM summaries often miss key insights even with full transcripts.
Concerns about LLM accuracy and overreliance
- Experiences of wrong factual answers (e.g., business hours) reduce trust.
- Distinction made between summarizing static documents vs answering ephemeral factual queries, but users note the UI doesn’t clearly signal this.
- Philosophical worries: outsourcing “what’s worth my time” and knowledge acquisition to opaque systems; risk of everything converging to the middle of the bell curve.
Reactions to the talk
- Several see it as a polished high‑level overview, potentially aimed at undergrads, with strong emphasis on Google’s own work and products.
- Others criticize it as marketing‑heavy and light on theory, architectural details, or genuine explanation of “why it works.”
- Some note relative underemphasis on ChatGPT‑style breakthroughs and frame differences between labs as execution, not research direction.
State and direction of ML research
- One view: current neural‑net + massive compute paradigm is plateauing; new, less data/compute‑hungry approaches will be needed.
- Another view: Transformers are becoming a foundational “MOS transistor” of ML, likely to dominate across modalities, with future work focusing on efficiency (e.g., reduced quadratic costs, database‑like attention, quantization) rather than new architectures.
- Skeptical commenters argue ML research is heavily driven by corporate incentives, dominated by computer vision and information‑retrieval use cases (ads, short‑form video, knowledge panels).
Business and competition
- Some suggest it’s a “wait and see” phase for AI/AGI startups: incumbents form an oligopoly, promising startups are likely acquisition targets.
- Others think Google will stay committed and that traditional keyword web search will diminish over time.