Race Dynamics And Safety Coordination Failure

Issue 89 Edition 2026-03-30 7 min read

General

Sources: 1 • Confidence: Medium • Updated: 2026-03-31 04:42

Key takeaways

Mallaby says ChatGPT's release made the AI competition feel like a war to Hassabis, who described it as rivals having 'parked the tanks on our front lawn.'
DeepMind was acquired by Google in 2014.
Mallaby says Hassabis initially believed language models were insufficient for general intelligence due to lack of grounding, but GPT-3's 2020 performance forced a reassessment.
AI progress is framed as a shift from symbolic rule-based methods to inductive, data-driven deep learning.
Mallaby says he had extensive biography access to Hassabis (over 30 hours of one-on-one interviews) on the condition that Hassabis could veto direct quotes while Mallaby could still use the underlying information.

Mallaby says ChatGPT's release made the AI competition feel like a war to Hassabis, who described it as rivals having 'parked the tanks on our front lawn.'
Mallaby reports that DeepMind and Anthropic were reluctant to release chatbot-style systems due to toxicity and hallucination concerns, while OpenAI released ChatGPT with fewer inhibitions, which rivals viewed as opportunistic and unfair.
The competitive dynamics of the AI industry are described as making rapid capability releases and acceleration difficult for any single actor to avoid even if they fear catastrophic risks.
Mallaby says Hassabis pursued safety theories including a 'singleton' scenario in which one lab develops AGI on behalf of humanity and ensures safety before release, but that ChatGPT undermined this premise by intensifying multi-actor competition.
Mallaby says the singleton safety idea became implausible once OpenAI was founded in 2015.
Mallaby says Microsoft publicly celebrated ChatGPT's lead over Google/DeepMind, including Nadella's 'can they dance' remark, which intensified hostility and made the competition a 'bare-knuckle fight.'

DeepMind was acquired by Google in 2014.
Mallaby says DeepMind's 2014 sale to Google was driven by rising capital needs for compute and talent and by Hassabis wanting to spend less time fundraising and more time on science.
Mallaby says Hassabis rejected acquisition interest from Facebook and Elon Musk and concluded Google would be the best parent due to resources and fit.

Mallaby says Hassabis initially believed language models were insufficient for general intelligence due to lack of grounding, but GPT-3's 2020 performance forced a reassessment.
Mallaby says that after the 2017 Transformer breakthrough from Google Research, OpenAI pivoted quickly to apply Transformers to language while DeepMind deprioritized language and lacked a team positioned to exploit the approach.
Mallaby describes a narrative arc in which OpenAI became the leading lab after ChatGPT's release, but by late 2025 he believes Google DeepMind's Gemini 3 was judged better than OpenAI's model.

AI progress is framed as a shift from symbolic rule-based methods to inductive, data-driven deep learning.
Neural networks are described as learning categories through trial-and-error parameter updates over labeled examples rather than explicit human-written rules, often yielding internal solutions that are hard for programmers to interpret.

Mallaby says he had extensive biography access to Hassabis (over 30 hours of one-on-one interviews) on the condition that Hassabis could veto direct quotes while Mallaby could still use the underlying information.
Mallaby says he used large language models such as Gemini to preprocess scientists' papers and generate comparative briefings before interviews.

What concrete, independently checkable evaluations support the claim that Gemini 3 was judged better than OpenAI's model in late 2025 (benchmarks, red-teaming results, enterprise adoption, or third-party rankings)?
What were the actual internal resource allocations and organizational decisions at DeepMind regarding language models after the 2017 Transformer breakthrough (team sizes, mandates, and timelines)?
How did the safety review and release gating processes differ across DeepMind, Anthropic, and OpenAI at the time of ChatGPT's release, and what incident rates followed?
What specific compute and talent cost trajectories or bottlenecks drove DeepMind toward acquisition in 2014 (budgets, projected training costs, hiring constraints)?
What measurable indicators show that AI safety concerns have receded from public discourse, and how do those indicators correlate with geopolitical/commercial pressures versus other factors?

Frontier AI competition is framed as warlike post ChatGPT, implying faster release cycles and reduced willingness to coordinate on safety gating, which could shift advantage to actors with distribution and deployment speed.
Compute and talent scale are portrayed as decisive, with DeepMind’s 2014 Google acquisition linked to capital constraints, implying consolidation toward hyperscaler backed labs and infrastructure providers.
Execution and prioritization may outweigh research origin, with a narrative that OpenAI pivoted faster to language while DeepMind deprioritized until GPT 3, implying org structure and resource allocation choices drive competitive positioning.

Publicly verifiable third party evaluations showing late 2025 model leadership claims, including benchmarks, red teaming outcomes, enterprise adoption, or independent rankings referenced in the unknowns.
Disclosed or credibly reported internal resourcing and mandates for language models after the 2017 Transformer breakthrough, including team sizes, timelines, and shifts after GPT 3 performance.
Observable divergence in safety review and release gating across major labs, paired with measurable incident rates post release that would indicate competition altering safety posture.

Lack of any independently checkable evidence for comparative model leadership claims, or contradictory third party evaluations that do not support the late 2025 judgment referenced.
Evidence that DeepMind did not face meaningful compute or talent constraints around 2014, or that acquisition rationale was unrelated to scaling needs, undermining the capital constraints narrative.
Evidence that safety concerns have not receded in public discourse or that gating practices and incident rates do not correlate with competitive pressures, weakening the coordination failure framing.