Lectura

Cohere North 2 adds token caps.

Its North Admin console tracks token use by user and agent, with tiers and alerts.

SiliconANGLE said Cohere framed this as a cost fix; no independent data.

Iterate.ai launched Lifeboat, an inference engine claiming to support 2,048 concurrent AI agent sessions on a single Nvidia RTX PRO 6000 GPU.

The engine boosts effective capacity by two to six times through key-value cache optimization and fair scheduling. It addresses memory overflow in long-context tasks by loading only necessary experts from mixture-of-experts models while keeping weights at full precision.

Each session runs in an isolated security capsule with hardware attestation for confidential computing. A free developer license is available, with the top-tier Confidential Computing edition priced at $499.99 per month.

Cohere and PricewaterhouseCoopers (PwC) announced a global alliance on October 5, launching first in Canada.

The partnership combines Cohere’s North agentic platform, enterprise models, and retrieval capabilities with PwC’s expertise in risk, regulation, and technology transformation. The goal is to help organizations identify high-value use cases and securely connect AI to trusted enterprise data.

The collaboration supports private cloud, on-premises, and air-gapped deployments to meet strict data sovereignty requirements in sectors like finance and public sector. No specific financial terms or initial client names were disclosed.

Safeworld, founded by Dr. Ding Zhao of Carnegie Mellon University's Safe AI lab, has emerged from stealth with a seed round exceeding $12 million.

Led by Shine Capital and a16z Speedrun, the company addresses the unpredictability of generative AI-controlled robots. It builds simulation environments populated with realistic digital humans to test edge cases at scale before physical deployment.

While traditional algorithms are predictable, GenAI-driven robots pose new safety challenges in unstructured environments. Safeworld offers third-party validation to help manufacturers prove system safety and build trust. Early partners include Gritt Robotics.

Aleph Alpha released Kolibri, an open-weight language model using a mixture-of-experts architecture with 78 billion total parameters and approximately 3 billion active per token.

German accounts for 21.3% of the training data, and the model was trained on 768 B200 GPUs in Germany and Finland. The company states it complies with the EU AI Act and targets public administration, aviation, and industry sectors.

Weights are available under the Apache 2.0 license on Hugging Face. Aleph Alpha reports a 71% score on German benchmarks and claims Pareto-optimal quality-cost performance, though these results remain vendor-self-reported and unverified by independent parties.

Huawei and Qualcomm announced on October 5 that they have entered into a broad, multi-year patent license agreement.

The deal includes cross-licensing of patent portfolios in areas such as 5G, computing, artificial intelligence, and networking. Additionally, Qualcomm will acquire certain US patents from Huawei in these fields.

Executives from both companies stated the agreement validates their respective innovations and foundational R&D contributions. Alan Fan, Huawei’s Chief IP Officer, noted it confirms Huawei's leading position in mobile communications.

The transaction is subject to regulatory approval. This follows a similar Wi-Fi patent cross-license agreement signed between Huawei and HP in August.

Getting machines to learn continuous actions by trial and error—how much to press the gas, how much to rotate a joint—is called reinforcement learning. The machine has to maintain an estimate throughout: how many points it can get by following the current way of acting. The 2018 paper TD3 found that this estimate tends to be biased upward, and the machine chases the inflated score; the fix is to maintain two independent scorers, trust only the lower one, and slow down changes in the way of acting.

Today's demos of robots learning to walk and grasp things are often still backed by this trial-and-error learning, so the problem of inflated scores is still there. TD3's "take the lower of two scorers" was not bypassed by later new methods; instead, it became a standard feature of mainstream algorithms. Any system that chooses actions by estimating scores should ask one question: has the overestimation been prevented.

If the actions are discrete—choose left or right, rather than how many degrees—then don't use it to judge; its validation is entirely on simulated robot tasks, so don't directly extrapolate the results to real robots or other tasks. And suppressing the scores has a cost: genuinely good opportunities are also suppressed along with it, and overestimation is merely replaced by underestimation.

《Addressing Function Approximation Error in Actor-Critic Methods》(2018)｜Next review 2027-09-20