Robots+5Oct 24
Image: The Threshold Report/GPT Image 2.5
Eka reports unfamiliar-object handling at 1,500 operations per hour
Eka Robotics says its robots can pick objects outside their training examples at roughly 1,500 operations per hour. Its model combines visual information with forces felt during contact and learns through reinforcement learning in simulation, meaning repeated simulated attempts guided by rewards. That approach could reduce the need to collect human demonstrations for each new factory task if the results transfer to working environments. Eka says the robots can also screw in a light bulb and insert wires.
Robots+5Oct 23
Image: The Threshold Report/GPT Image 2.5
SuperMimic reports over 1,000-fold lower robot-training compute
After three hours on one graphics chip, SuperMimic reportedly outperformed a leading method for teaching robots to track motions. The method used over 1,000 times less total graphics-chip training time, potentially making robot-motion experiments possible without large chip clusters. Other research demonstrations included Reflex's learned whole-body reactions and an ETH Zurich four-legged robot solving a giant Rubik's Cube autonomously with its feet. A vision-guided KAIST humanoid pushed objects up to 35 kilograms and lifted up to 20 kilograms. SuperMimic failed on 25 of 129,514 benchmark motions, compared with 179 failures for the leading comparison method. A longer SuperMimic run of 9.8 hours on two graphics chips recorded zero failures.
Agents+2Text+2Oct 24
Image: The Threshold Report/GPT Image 2.5
Nemotron Bolt becomes downloadable for local tool-using assistants
Nemotron Bolt's downloadable model files support device-based assistants that use tools across a conversation, with commercial and non-commercial use allowed under OpenMDW-1.1. NVIDIA built NVIDIA-Nemotron-3.5-Bolt-E4B-4GB from Gemma-4-E4B and lists 8 billion total parameters (values learned during training), with 4 billion counted as effective. Input and output can each reach 64,000 tokens (units of text), and the files come in Q4_K_M (a compressed format). Applications can choose thinking or conversational modes to adjust response behavior against the time needed to answer, including for interactive personas and embedded interfaces. NVIDIA attributes improvements in following instructions, using supporting information, and handling tools over multiple exchanges to its additional training.
Agents+2Oct 23
Image: The Threshold Report/GPT Image 2.5
Claude Enterprise unifies chat and Cowork with continuing cloud sessions
A Claude Enterprise conversation can move into code execution and file work that keeps running in the cloud after a laptop closes. Anthropic combined chat and Cowork on web, desktop, and mobile, allowing employees to resume document, analysis, or presentation tasks from another device. Administrators get an Efficient mode limited to Sonnet and Haiku. Cloud sessions run in isolated virtual machines (software-separated computers), with outside network access off by default for Enterprise and connector credentials kept outside the agent's work environment. Anthropic says typical Opus 5.5 workloads cost 40% less than Opus 5, while Sonnet 5.5 costs up to 30% less than Sonnet 5 for most work.
Compute & Infra+2Oct 24
Image: datacenterdynamics.com
ZTE launches data-center module with on-site liquid-cooling upgrades
The ZXEDC M8000, launched by ZTE, lets operators convert racks from air to liquid cooling on site without replacing the whole data-center module. It combines racks, cooling distribution and piping, high-voltage direct-current power, and monitoring for mixed general-purpose and AI computing. ZTE says rack power ranges from 20–60 kilowatts, with peaks of 250 kilowatts. The company claims power usage effectiveness as low as 1.15, meaning total facility power is 1.15 times the power consumed by computing equipment. ZTE also claims up to 98% power-supply efficiency and UL system-level leak-proof certification. The architecture can accommodate three generations of IT hardware upgrades, according to ZTE.
Compute & Infra+2Oct 24
Image: datacenterdynamics.com
Crown Castle reports more than 100 deployed edge-computing trial sites
More than 100 sites are deployed in Crown Castle trials of edge computing, meaning computing located close to users. Several trials are running concurrently, including work supporting Available Infrastructure's network for AI computing near customers. Typical installations use 0.2–0.5 megawatts and reuse existing tower-site structures, power, and broadband connections.
Text+2Oct 23
Image: The Threshold Report/GPT Image 2.5
LitPivot improves literature grounding in a small research-proposal study
In a 17-researcher comparison, Ai2 and University of Pennsylvania collaborators found that LitPivot helped produce proposals better grounded in relevant literature than a text editor and an AI chat tool. LitPivot gathers and groups papers around a proposal's problem, contribution, or evaluation plan, attaching supporting excerpts to critiques and suggested revisions. Researchers can change the evidence, accept or reject edits, and see when a revision leaves another section inconsistent. Participants also rated their proposals as more novel and useful, but those judgments were their own assessments. Both conditions used the same paper collection, and two subject experts rated the proposals without knowing which tool produced them.
