Import AI 479: Spacewar; low-latency tasks versus LLMs; Genie 4

1. Key Themes

Theme: Real-time latency is a major gap in frontier AI agents

LLMs collapse on tasks where time passes while they think

Realtime-Bench shows that models do far worse when deliberation costs real time. The benchmark is designed so that "delayed actions directly affect success."

"GPT-6 Astra (high) got the highest pass rate on the 'wall clock' evaluation setting, with 36%. By comparison, with a paused clock, models with high reasoning get closer to 70% or even 80% success."

The gap between paused-clock and wall-clock performance suggests the underlying capability exists but cannot be deployed under time pressure.

The benchmark decomposes the failure into three distinct skills

The test isolates three separable capabilities, which gives builders a roadmap for where to improve.

"temporal state awareness ('RT-Perceive'), where models can use observations to track changes in the world; adaptive deliberation ('RT-Deliberate'), where models alter their reasoning length and depth according to available time; and time-aware action selection ('RT-Act') where models need to appropriately time their actions for the situation."

Low scores map to where agents can't yet be trusted

The author frames the benchmark as a guide to deployment risk.

"This benchmark is an effective proxy for where AI agents will break - the low scores here point to tasks where AI agents cannot yet be expected to perform well."

The task mix includes commercial workflows, so the gap matters for agentic commerce: "30 web tasks in four categories (travel, e-commerce, service, and security)."

Theme: Space is becoming a contested strategic domain, and AI infrastructure is moving there

Adversaries are building a broad offensive counterspace arsenal

The DIA report describes kinetic, directed-energy, jamming, cyber, and nuclear threats to satellites.

"China and Russia are developing a formidable spectrum of offensive capabilities to degrade or destroy space services of other countries, from kinetic destruction to sophisticated jamming and cyber operations."

"By the late 2020s, China may field space-based directed energy systems and higher-powered ground-based directed energy systems capable of threatening satellites in higher orbits."

China's launch cadence is a leading indicator of capability

"China has more than tripled its space launches per year in the past decade, with 22 launches in 2016 and 93 launches in 2025."

Satellite-on-satellite "shadowing" is already happening

"In 2019, Russia launched satellite Cosmos 2542, which released a second satellite, Cosmos 2543. Cosmos 2543 then began following a U.S. national security satellite and approached it multiple times at a distance close enough to create potentially dangerous operating conditions."

Nuclear detonation in space threatens every orbital layer

"A nuclear detonation in space could be tailored to impact satellites in most orbits—LEO for commercial and reconnaissance satellites, GEO for critical communications, and MEO for navigation and timing satellites."

Space-based compute raises the stakes

Clark ties this to AI infrastructure investment.

"As people spend more resources on building space-based data centers (and eventually, perhaps, lunar-based ones), space will become ever more strategically and commercially valuable and ever more contested."

For investors, this implies orbital assets carry geopolitical risk that terrestrial data centers do not.

Theme: A neuro-symbolic revival is being pitched as the path to reliable reasoning

Metis argues LLMs leave reasoning structure implicit

The startup claims modern architectures lack two things that made AlphaGo work.

"The first missing element is an explicit and persistent epistemic state… the second missing element is a criterion that allows the system to choose reasoning moves that make progress towards the goal."

Provenance and auditability are the commercial hook

The design requires evidence for every accepted claim, which targets hallucination and enterprise trust.

"Model-generated content cannot enter the accepted epistemic state merely by being plausible; it must pass warrant assessment, which lowers hallucinations. Sources of evidence can be attached directly to the propositions they support, providing a clear chain of provenance. The reasoning process becomes explicitly auditable."

Framing: two incomplete successes

"While the history of AI can be understood as a contest between symbolic and neural methods, it may be more constructive to view it as a story of two incomplete successes."

Caveat: Clark notes "This is more of a position paper than a technical paper, so they don't have a ton of detail here," so there is no empirical validation yet.

Theme: World models are advancing toward interactive, agent-guided simulation

Genie 4 adds real-time editing, physics, audio, and a planning agent

Google DeepMind's update bundles several capabilities that make generated worlds more usable.

"You can use natural language to modify a simulation in real-time — like changing the weather, moving the camera, or introducing a new person or object."

"Genie 4 is our first real-time world model to generate audio natively."

Pairing a fast LLM with a world model improves causality

"Google has embedded Gemini 3.8 Flash into Genie 4, so that the fast Flash model 'serves as a guiding agent in the loop', providing continual calibration. 'Gemini continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded.'"

Still early, but pointing at infinite generated content and possibly data

"Though the technology is currently quite janky (e.g., robotic voices, a dreamlike sort of inconsistency in the visualized environments, etc), it points towards a world where we can generate infinite forms of entertainment (and potentially data?) via creating and prompting custom worlds."

Distribution is gated: "rolling out access to Genie 4 currently as a research preview for some trusted testers and will eventually come to Google AI Ultra subscribers."

2. Contrarian Perspectives

Perspective: The more "human-like" architecture is the messy LLM, not the rigorous symbolic system

Conventional framing treats explicit, logical reasoning as the more human-like ideal. Clark argues the opposite: structured epistemic systems are the alien ones.

"Humans do not reason from some pristine base of facts, rather it's all a big messy blob of memory and intuitions and some rules... In this way, contemporary LLMs are more like humans than the system described here."

Implication: reliability-oriented architectures may sacrifice the very qualities that make LLMs flexible, and may reveal "the differences between our own intelligence and that of machines."

Perspective: Reasoning capability is not the bottleneck for agents; time-awareness is

The consensus assumes smarter models will unlock agentic use cases. The Realtime-Bench data suggests capability and deployability are different problems.

"With a paused clock, models with high reasoning get closer to 70% or even 80% success" versus 36% at best under wall-clock conditions.

The best model loses roughly half its performance simply because thinking takes time. Note that the benchmark was "designed so that consuming more tokens delays execution regardless of actual inference latency" in its token-clock mode, so more reasoning can actively hurt.

3. Companies Identified

Metis

Description: Startup building AI architectures that fuse symbolic and neural reasoning. Why mentioned: Pitching an explicit "epistemic state" and auditable reasoning core, aiming to replicate AlphaGo-style reliability for open-domain problems. Quotes:

"Metis wants to do for modern open-domain problems what DeepMind did for closed-domain problems like board games via its work on AlphaGo and AlphaGo Zero."

Google DeepMind

Description: Google's AI research lab; creator of the Genie world-model series. Why mentioned: Released Genie 4 with real-time modification, physics, native audio, and an embedded Gemini agent. Also referenced for AlphaGo. Quotes:

"Genie 4 is our first real-time world model to generate audio natively."

Google (Gemini models)

Description: Maker of Gemini 3.8 Flash, Gemini 3.1 Flash Live, and Gemini 3.8 Live. Why mentioned: Gemini 3.8 Flash is embedded in Genie 4 as the guiding agent, and Gemini models were evaluated on Realtime-Bench. Quotes:

"Google has embedded Gemini 3.8 Flash into Genie 4, so that the fast Flash model 'serves as a guiding agent in the loop', providing continual calibration."

OpenAI

Description: Frontier lab. Why mentioned: GPT-6-Astra topped the Realtime-Bench wall-clock results; other OpenAI models (GPT-5.6 Sol, GPT-Realtime-2, OpenAI Decisions) were also tested. Quotes:

"GPT-6 Astra (high) got the highest pass rate on the 'wall clock' evaluation setting, with 36%."

Anthropic

Description: Frontier lab. Why mentioned: Claude Fable 5.1 and Claude Opus 5 were benchmarked on Realtime-Bench. Quotes:

"The authors tested out models from OpenAI... Anthropic (Claude Fable 5.1, Claude Opus 5), Google..."

Deepseek and Moonshot AI

Description: Chinese AI labs. Why mentioned: Deepseek V4.1 and Kimi-K3 were included as Chinese comparison models. Quotes:

"...as well as Chinese models from Deepseek (V4.1) and Moonshot AI (Kimi-K3)."

NVIDIA and Steadyworks

Description: Corporate collaborators on Realtime-Bench (alongside universities). Why mentioned: Part of the author group, a signal of industry interest in real-time agent evaluation. Quotes:

"The benchmark was built by people at the University of Texas at Austin, Ohio State University, University of Illinois at Urbana-Champaign, NVIDIA, Northwestern University, and Steadyworks."

4. People Identified

Jack Clark

Description: Author of Import AI. Why mentioned: Curator and commentator throughout; offers the framing that these benchmarks and reports map where AI and its infrastructure are heading. Quotes:

"This benchmark is an effective proxy for where AI agents will break."

Iain M. Banks (inspiration reference)

Description: Sci-fi author of the Culture series. Why mentioned: Cited as an inspiration for the closing fiction piece, along with Hermann Hesse. Quotes:

"Things that inspired this story: Player of Games by Iain M Banks; Simcity; game playing; urban design; The Glass Bead Game by Hermann Hesse."

5. Operating Insights

Design agent workflows around latency budgets, not just accuracy

If you're building agents for time-sensitive domains (ticketing, auctions, trading, gaming, security), test under wall-clock conditions rather than idealized ones. The benchmark's own example shows what to look for:

"An agent purchasing a ticket must track the sale status and assess whether observed availability is still current (RT-Perceive), limit further deliberation when the purchase window is closing (RT-Deliberate), and choose between waiting for sales to open and purchasing while tickets remain available (RT-Act)."

Practical implication: build in explicit mechanisms for time-bounded reasoning (e.g., caps on thinking length when a window is closing) and stale-state checks.

Use a fast model as an in-the-loop controller alongside a heavier generative system

Genie 4's architecture is a reusable pattern: a small, fast model continuously monitors outputs and replans.

"Gemini continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded. As the world shifts or the user intervenes, Gemini adapts its plan in the moment."

This is a template for pairing slow, high-quality models with fast supervisors in latency-sensitive products.

Build provenance and verification into reasoning pipelines for trust-sensitive customers

Metis's design principle is useful even without adopting its architecture: require evidence before state changes.

"Any change to a claim's evidential status must be licensed by evidence or verified through explicit calculation and approved by an independent judgment system."

For enterprise, legal, or medical products, attaching sources to claims and gating acceptance on verification can differentiate on auditability.

6. Overlooked Insights

Realtime-Bench's "token clock" quietly penalizes verbose reasoning

The benchmark includes a condition that "advances the environment at a fixed rate per generation token," so "consuming more tokens delays execution regardless of actual inference latency." This means the benchmark can measure reasoning efficiency independent of hardware speed, which is a way to compare models' deliberation economy that could matter more as reasoning-heavy models proliferate.

Russian satellites are being launched into isolated orbits for unexplained testing

The DIA notes a February 2022 launch "into an unusual and isolated orbit not used by any other spacecraft," with a stated purpose (testing space radiation effects) that the report treats skeptically. Alongside China's spaceplane that "released and interacted with an object" in 2023, this points to growing on-orbit rendezvous and proximity capability, a niche with potential defense and space-situational-awareness demand.