Fire at New York's $3.2 Billion AI Data Center Exposes Emergency Gaps

01Fire at New York’s $3.2 Billion AI Data Center Exposes Emergency Gaps as Operator and Fire Department Differ on Fixes

TeraWulf owns and operates Lake Mariner, the $3.2 billion AI campus in Somerset, New York, and has confirmed that it is responsible for emergency preparation there. The company says it implemented fixes after an early-June fire, but the local fire chief's mid-August account raises questions about whether the promised access equipment and water supplies were in place.

Firefighters entered a still-unfinished building at the Lake Mariner data center in early June without knowing what chemicals were producing the heavy black smoke inside. Steve Matisz, chief of the Barker Fire Department, said his crew went in “kind of blind” because the safety sheets meant to identify the hazards had apparently burned in the fire.

The blaze exposed gaps in emergency preparation at the $3.2 billion AI computing campus in Somerset, New York. Firefighters reportedly found no working alarm or fire-suppression system, while three hydrants at the site could not supply water. The safety documents that the department was legally entitled to review were also reportedly destroyed in the building.

Lake Mariner, being developed at a former coal mine on Lake Ontario, is one of New York’s largest AI data center projects. Its structure divides ownership, financing, operation and computing demand among several companies, making responsibility less obvious to local responders.

TeraWulf owns and operates the data center on land leased from a company owned by its own chief executive. Fluidstack, a UK-based AI company, is due to run the facility. Google has agreed to guarantee Fluidstack’s lease payments and holds warrants that could give it a future 14 percent stake, while Anthropic is among the AI companies whose computing demand the project is intended to serve.

TeraWulf confirmed that it—not those other participants—is responsible for operational safety and emergency preparedness across the campus, including required safety systems and coordination with local first responders.

After an incident review with the county, TeraWulf chief strategy officer Kerri Langlais said the company had implemented Knox key boxes, added hydrants and prepared portable emergency packs containing safety data sheets. The boxes are intended to give emergency personnel access to secured facilities, while the document packs would keep hazard information available during an incident.

Matisz’s account in mid-August suggested that at least some of that work remained incomplete. He said a Knox box representative had visited Lake Mariner and a meeting was scheduled to build out the program. He was firmer about the water supply: as far as he knew, the hydrants were still dry, and he had not seen work performed on them.

The available account does not establish the fire’s cause, whether anyone was injured, or whether regulators determined that any law was broken. It also does not confirm the hydrants’ status after mid-August.

02Iris unveils two search agents trained by alternating supervision and live-search reinforcement learning

The Iris team has unveiled Iris-mini and Iris-pro, two search agents designed for complex web tasks that require several retrieval steps rather than a single query. Iris-mini is the smaller model and Iris-pro the larger one. Both were trained through a pipeline that alternates supervised fine-tuning, or SFT, with reinforcement learning against live search. The team plans to release the weights and its complete data, training and evaluation recipe, but gave no release date or independent production results.

To prevent training questions from being solved through simple keyword matching, the team worked backward from the hyperlink structure of a web corpus. Starting with a seed page and its outgoing links, it created an entity graph and authored questions around multi-hop chains within that graph.

Every entity other than the answer was then rewritten as a descriptive reference, removing exact names that could reveal the required search path through string matching. A question entered the training set only if a reference model failed to answer it without external material but succeeded when given the supporting evidence.

The resulting questions were converted into search trajectories. Iris filtered these examples at both the complete-trajectory and individual-turn levels before using them for SFT. It then optimized the policy through reinforcement learning with live search; the reward judge and observation summarizer ran inside the training cluster, while overlong search rollouts could be interrupted and resumed from their committed prefix.

The team calls its alternating process “SFT-RL climbing.” After each reinforcement-learning round, the hardest successfully completed and most efficient trajectories were returned to the next supervised-training pass, creating a cycle in which live-search successes supplied increasingly demanding demonstrations.

The available source reports benchmark results but does not define the tasks or scoring scales. With inference-time context management enabled, Iris-mini scored 82.2 on BrowseComp, 84.8 on BrowseComp-ZH, 86.9 on DeepSearchQA and 52.3 on HLE. Iris-pro scored 88.6, 85.1, 92.9 and 56.4 on the same benchmarks. The team describes these as the strongest overall results among open-source search agents in their respective parameter ranges.

Because context management can materially affect these benchmarks, the team evaluated each test both with and without it while keeping the tools, context limit and judge fixed. All reported results came from a single agent, without sub-agents or test-time verification.

03Motion-Omni Lets Dialogue Models Generate Speech and Full-Body Motion Together at Real-Time Speed

Digital avatars typically rely on a two-stage pipeline to speak and move: a dialogue model first generates the complete audio response, then a separate co-speech motion model turns that finished audio into gestures. That second inference pass adds latency, while the separation prevents the speech and movement systems from being optimized together and can leave actions out of sync with the voice.

Motion-Omni replaces that cascade with an end-to-end framework that jointly produces spoken dialogue and accompanying motion. According to the researchers, the system generates facial expressions and hand, upper-body and lower-body movements directly from the same hidden states used to produce speech, allowing the outputs to be created together rather than one after the other.

Experiments described in the Motion-Omni paper indicate that sharing an architecture is not enough by itself. When the researchers froze the speech pathway and trained only the motion module, the resulting movements remained misaligned with the audio. Alignment recovered when they jointly adjusted the underlying large language model, speech generator and motion generator under both speech and motion objectives, while retaining the model’s spoken-dialogue ability.

To create training supervision at scale, the team generated consistent-voice spoken responses and used a replaceable motion “teacher”—an existing model that supplies training targets—to assign pseudo-labels. That pipeline produced 422,856 quality-ranked speech-and-motion pairs covering 1,402 hours.

The resulting Motion-Omni-Q7 model, built on Qwen2.5-7B-Instruct, came within 2% of the same-audio teacher cascade on reference-free motion metrics. That comparison applies specifically to those motion measures, rather than every aspect of movement quality. The model also surpassed all evaluated non-teacher cascades on beat correlation and diversity.

Motion-Omni-Q7 ran 5.4 times faster than the teacher cascade and recorded a real-time factor of 0.78, meaning inference took less time than the generated output lasted. Its word error rate was 2.62%, the lowest among the omni-modal systems in the comparison.

The researchers also released SwDA-500 and a public evaluation protocol for open-ended full-body spoken dialogue. The source does not establish performance in deployed products, long conversations, different characters or scenes, and does not specify the hardware used for the reported speed.