Andrew Dai, a former Google Brain and DeepMind researcher and co-founder of Elorian, argues that frontier AI remains unreliable at detailed visual tasks such as counting objects, tracking robot actions and interpreting complex designs. The practical point is less glamorous than another benchmark victory: systems that recognise a chessboard may still fail when asked how many squares are visible on it.
Watch Desk analysis
What happened
In an appearance on the AI Engineer podcast, as reported by Finance BigGo, Dai drew a line between rapid visual recognition and sustained visual reasoning. He said current models can identify familiar objects quickly but often hallucinate a complete scene when a task requires careful inspection.
Dai gave examples involving partial chessboards, board-game pieces and robot-arm videos. In one account, a model answered 32 when asked to count the white squares visible in an incomplete chessboard image, apparently recalling the standard number of white squares on a full board rather than inspecting the image. He also said models missed important actions in a video of a robot handling a cup and cooker, including lifting a lid and turning on a stove.
Dai’s argument is that this weakness matters most in fields where mistakes are physical or expensive: robotics, construction, mechanical design and safety compliance. He said Elorian is building visual-reasoning systems aimed at those sectors, with an API for robotics planned by the end of 2026.
Why it matters
The distinction between recognising an object and reasoning about its exact position, count or movement is useful because deployment depends on the latter. A model that labels a construction site correctly but misses a helmet, a hazard or a change in machinery is not merely having an off day. It is failing at the bit people hoped would make computer vision operationally useful.
Dai also challenged the way progress is measured. He argued that small-image benchmarks such as ARC-AGI do not capture the complexity of real-world visual work, while tests of geometric alignment, spatial intelligence and object permanence remain underdeveloped. That is a serious criticism, although it comes from a founder whose company is selling an alternative approach.
Our read
Dai’s examples make a persuasive case that visual AI needs tougher, more realistic tests. They do not prove that every frontier model fails in the same way, nor that Elorian’s proposed visual chain-of-thought approach will solve the problem. The podcast account contains no independent comparative evaluation of Elorian’s system, and its promised API has not yet shipped.
Still, this is a worthwhile shift in the argument. AI is often judged by whether it can produce a plausible answer. In robotics and engineering, the more important question is whether it noticed the right thing before acting. Plausibility is a charming quality in a dinner guest. It is less charming in a machine operating a stove.
What to watch
- Whether Elorian delivers its planned robotics API by the end of 2026.
- Independent tests of visual counting, spatial reasoning and long-video consistency.
- Whether new benchmarks measure real-world geometry and object permanence rather than tiny-image pattern matching.
- Evidence that Elorian’s visual data and architecture produce better results than existing multimodal systems.
Discussion spark: Should AI companies be judged less by broad benchmark scores and more by difficult visual tests that reflect the physical work they want to automate?
Sources and evidence
- Andrew Dai Says Frontier AI Can't Count Chessboard Squares – and That's a $100 Billion Problem – finance.biggo.com (23 September 2026, 16:08 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.