June 17 – AI Research

Reframe the Question: What Survived the Noise in AI

June 17 delivered the usual flood of AI commentary, yet only a handful of posts from researchers cleared the bar by carrying actual claims, positions, or named references instead of bare links, single-word reactions, or offhand pointers. Those that did offered a through-line of clarity over added layers of hype or complexity.

François Chollet put the principle plainly: the hardest problems are rarely solved by adding more complexity to the solution. They yield instead when the question itself is reframed until a simpler, clearer answer reveals itself. That same impulse to strip away noise and name what is actually happening ran through the day’s most substantive contributions.

Gary Marcus returned to the point repeatedly. He highlighted a new paper from Google DeepMind, the University of Waterloo, ANU, and UCL showing that even what the authors call “competent AGI” has not yet been achieved, let alone expert or superhuman levels. Hypesters may pretend otherwise by traditional standards of the term, Marcus wrote, but he fully concurred that claims to the contrary amount to marketing.

He aimed the same directness at corporate strategy. Addressing Mark Zuckerberg, Marcus stated that AGI cannot be reached through data labeling alone and described the conversion of Meta’s once top-notch AI research division into a data-labeling sweatshop as one of the dumbest blunders in corporate history.

When challenged on his own earlier analysis, Marcus responded with characteristic economy: “really? where have you been?” and pointed to his February 2024 post listing five trends that continue to occupy GenAI—the politics and inadequacy of guardrails, copyright litigation, customer retention, the lack of a moat as performance converges, and deepfakes—as one example among many he has made over time.

Even in shorter form, he observed that moral problems can be nibbled away one bite at a time. Together these posts modeled the very discipline Chollet described: cutting through accumulated complexity to isolate the real questions still facing the field. On June 17, at least a few voices chose that reframing over fresh layers of spin.