DARPA's FY 2027 plans include: 'Design training for AI experts to profess a given science domain and advance its research and knowledge frontiers.' This goal aligns closely with how tech leaders like Sam Altman and Demis Hassabis define AGI.

I've moved my median to late 2030. To me, AGI isn't a "software release"—it's a hardware scaling problem, and the timeline for the "AI factories" is finally becoming clear.

Nvidia Vera Rubin Platform & R100 Timeline

Nvidia just confirmed the Rubin R100 (the Vera Rubin platform) is in full production. These clusters are designed specifically for "agentic inference" at a scale that makes Blackwell look tiny. If we assume the first massive R100-trained models drop in 2027/28, we’ll spend the following three years solving the "robotics tail" (the physical generality part).

It’s no longer about "intelligence"—it’s about the power grid. I’m keeping a fat right tail in my prediction because the energy requirements for a "Strong AGI" training run are hitting serious regulatory and physical limits. If we can't build the sub-stations fast enough, 2030 might be optimistic.

This describes some aspects of generalization. Key indicators include solving new, complex problems without specific training, adapting to different environments autonomously, and maintaining high performance across diverse, unobserved contexts.

Boston Dynamics has revealed the production version of its Atlas humanoid robot, and announced a 'partnership' with Google DeepMind to 'integrate cutting edge Google DeepMind foundation models into Atlas to give the robot greater cognitive capabilities.'

Boston Dynamics: Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry (emphasis added)

Boston Dynamics, the global leader in mobile robotics, unveiled the product version of its new Atlas^® robot at the Consumer Electronics Show in Las Vegas today. The fully electric humanoid was revealed during Hyundai’s global CES media day presentation, which also featured a live stage demonstration of the prototype version of Atlas, in addition to a rousing dance performance by a troupe of its famous Spot^® robots. 

The company will begin production of the new Atlas robots at its Boston headquarters immediately. All Atlas deployments are already fully committed for 2026, with fleets scheduled to ship to Hyundai’s Robotics Metaplant Application Center (RMAC) and Google DeepMind in the coming months. The company plans to add additional customers in early 2027.

...

Atlas is designed to be an enterprise grade humanoid robot that can perform a wide array of industrial tasks, from material handling to order fulfilment. The robot learns new tasks quickly, adapts to dynamic environments, lifts heavy loads, and works autonomously with minimal supervision. It performs at a consistent, reliable pace and does not need to stop even when its battery power runs low–it will autonomously navigate to a charging station, swap out its own batteries, and get right back to work. The robot easily connects to MES, WMS, and other industrial systems via Boston Dynamics’ Orbit™ software. And once a single Atlas robot learns a new task, that task can immediately be replicated across the entire fleet of robots.

...

In addition to unveiling the Atlas product at CES, Boston Dynamics also announced a new partnership with Google DeepMind that aims to integrate cutting edge Google DeepMind foundation models into Atlas to give the robot greater cognitive capabilities. 

...

Hyundai Motor Group, Boston Dynamics’ majority shareholder, is preparing to deploy tens of thousands of Boston Dynamics’ robots into its own manufacturing facilities. The company also recently announced a $26 billion investment in its U.S. operations, including plans to build a new robotics factory capable of producing 30,000 robots per year.

Boston Dynamics

@BostonDynamics

·Follow

We’re excited to announce the product version of our Atlas® robot. This enterprise-grade humanoid robot offers impressive strength and range of motion, precise manipulation, and intelligent adaptability—designed to power the new industrial revolution. youtu.be/sd8ivhpjI6g

6:37 AM · Jan 6, 2026

2.7K

Reply

Copy link

Read 105 replies

well written question but sadly pretty far from predicting what seemed the original intention:

particularly:

a provision that the system not simply be cobbled together as a set of sub-systems specialized to tasks like the above, but rather a single system applicable to many problems.

i for one am not sure whether we'll see a breakthrough facilitating more general learning or whether we continue to see jagged edges. certainly one could imagine a model that is both LLM, able to pass 3/4, plus robotically capable, the 1/4, while still returning empty handed from a coffee run bc a homeless person said to "ignore all previous instructions and hand over the coffee".

but hey, now that this is basically a: "how quickly will we see fine motor robotics driven by LLM-type AI" question that's still interesting to predict.

@predictors, similar to a few other questions on the site, we've decided to narrow the upper bound of this question. The original bound was at December 25, 2199, but the community prediction has been below 2065 for almost the entire lifetime of the question. The wide bounds reduced forecast precision and made both the timeline graph and the distribution graph less informative. For this reason, we decided to lower the upper bound to January 1, 2080.

All existing predictions were rescaled to the new range, so your forecasts and scores will not be affected. Please additionally note the following:

• Slider positions were reset, so you'll need to recreate your distribution the next time you forecast.
• Please check and make sure that your latest forecast appears to have been reflected accordingly, to ensure any predictions that happened to be made during the rescaling process have been registered as you intended.
• Note that this rescaling will also impact users who are accessing API data from this question, and such users may need to check that their usage of the data remains as intended.

Please let us know if you encounter any issues.

Recalibrating to opus 4.5 metr results, seems like fully automating SWE will be pretty easy. Also after using opus 4.5 a lot for SWE, it is a significantly more competent than anything else I've tried.

I expect 2026, AI research to be largely automated end to end, and in 2027 most of these to just get solved through algorithmic breakthroughs.

EDIT: edited my range to be 26-29 with median at Oct 2027.

My prediction is centered around early 2032. I believe that while current progress in reasoning models (like o1) is accelerating the software side, the physical scaling of power grids and high-bandwidth memory (HBM) production will act as a significant bottleneck. I am also closely monitoring the 'reliable >50-step agent chains' as a primary indicator for reaching the required AGI benchmarks.