Kosmos AI Scientist posts 79.4% accuracy, 1,500‑paper runs – Google tests Co‑Scientist

Edison Scientific unveiled Kosmos, an autonomous research system that can synthesize ~1,500 papers and write ~42,000 lines of analysis code in a single run, with 79.4% conclusion accuracy and full traceability to code and citations Altman endorsement, Launch article. The team highlights seven example discoveries and a structured world‑model approach that lets the agent stay on‑objective over millions of tokens.

• Beta users report a single 20‑step run replaced about 6.14 months of expert work, with perceived work scaling linearly with run depth scaling chart.

Why this matters: Kosmos packages long‑horizon research into repeatable, auditable workflows. That’s the piece lab leads and R&D heads need to justify compute and compliance at the same time.

Internal strings and demos show Google piloting two multi‑agent flows inside Gemini Enterprise: Idea Generation and a Co‑Scientist that, per run, spends ~40 minutes to generate and tournament‑rank ~100 ideas against user‑set criteria feature leak, Feature brief. The 3‑step loop takes a research goal + data, spawns specialist agents to explore, then evaluates and ranks based on an explicit rubric.

Why this matters: Teams get a repeatable front‑end for directed ideation with built‑in evaluation, which is the bottleneck for scaling literature triage and hypothesis pruning across orgs.

Google rolled out a Deep Research mode in NotebookLM that can autonomously browse hundreds of pages, synthesize findings into a structured report, and attach an annotated source list; it also expands supported source types (e.g., Drive URLs, Sheets, images) for mixed‑media research sets feature demo, Google blog post. Early user tests call it an “outstanding learning tool,” noting integrated mind maps, flashcards, and quizzes for follow‑up study hands‑on notes.

Why this matters: This is a ready‑to‑try research assistant with long‑running retrieval and auditable outputs—useful for product reviews, policy scans, and backgrounders that used to take days.

A comprehensive survey of scientific LLMs compiles 270 datasets and 190 benchmarks, proposes a taxonomy spanning raw observations→theory, and tracks a shift from single‑turn quizzes to process‑based grading of steps, tools, and intermediate results paper thread, ArXiv paper. The authors advocate closed‑loop agents that plan experiments, call simulators or labs, validate outcomes, and update shared knowledge—framing how to train and evaluate systems beyond static corpora.

Why this matters: It’s a roadmap for engineers stitching models, tools, and evaluators into credible pipelines for scientific work, with benchmarks that reward the process—not just the final answer.

At a UBS Q&A during GTC, Jensen Huang argued that customer ASICs can’t match NVIDIA’s full‑stack “AI factory” approach, citing an internal roadmap claiming up to ~40× beyond Hopper and the ability to place $100B‑scale POs with end‑to‑end systems and supply chain confidence transcript highlights. For infra leads, the message is clear: buyers will be sold on time‑to‑revenue, not chip lists.

This frames procurement around platform certainty and execution risk. If you’re modeling long‑lead data center bets, build scenarios where ASIC options don’t materially lower TCO once software, networking, power, and delivery timelines are included.

OpenAI’s observability team profiled node‑level Fluent Bit and found fstatat64 calls (triggered by inotify) burning ~35% CPU; turning that path off returned ~30,000 CPU cores to Kubernetes clusters processing nearly 10 PB/day of logs talk recap, with methodology and impact shared in the KubeCon session KubeCon talk. This is a big ops win: same workload, half the CPU.

If you run Fluent Bit, replicate the perf tracing, test inotify behavior under heavy appenders, and stage a rollout behind feature flags. Savings at this scale can fund more inference capacity immediately.

Groq lit up a 4.5MW data center in Sydney in partnership with Equinix Fabric, bringing low‑latency token serving to Australia and the wider APAC region launch note, with details in the company’s release press post. For teams in Australia, this cuts cross‑ocean latency and can lower per‑request costs when routing to closer endpoints.

Expect regional routing policies and capacity reservations to matter. If you’re piloting Groq, test latency deltas from Sydney versus US/EU regions and adjust traffic shaping accordingly.

Morgan Stanley exhibits circulating today show rental pricing for 8× H200 and early B200 nodes surging at launch, then stepping down as supply ramps—yet not returning to prior baselines chart thread. The takeaway for capacity planners: scarcity premiums ease, but structural demand keeps floor prices higher than last gen.

Model budgets around staged price relief, not a full reversion. Lock short terms for the peak window; renegotiate as additional capacity lands.

A widely shared Gamers Nexus breakdown reports DRAM pricing up ~3× in recent months, with knock‑on effects for NAND and GPU VRAM as AI servers absorb supply; prior oversupply cuts and potential manufacturer coordination are cited as drivers video note, echoed by community commentary flagging lab lock‑ins market note. This affects both server buildouts and on‑device edge AI plans.

YouTube analysis

Budget buffers for memory should widen. When speccing clusters or local inference nodes, watch lead times and consider pre‑buys on DIMMs/VRAM‑heavy SKUs before the next allocation bump.