Afternoon edition — Saturday 24 October 2026

NVIDIA published Nemotron 3.5 Bolt E4B 4GB on Hugging Face on 23 October, the date it had announced on 15 October alongside Nemotron 3.5 Super VL and Bolt 74B-A9B. The small model is post-trained from Google's Gemma-4-E4B-it for fast conversational agents with multi-turn tool use on GeForce RTX, Jetson and RTX or DGX Spark hardware, and it is released under the OpenMDW-1.1 licence.

MarkTechPost reports that NVIDIA skipped supervised fine-tuning and trained the model with reinforcement learning only, that its 4-bit build used 3.1 to 4.3 GB of video memory in NVIDIA's tests, and that it led 15 of 21 benchmarks against Gemma 4 E4B, Qwen3.5-4B and Nemotron 3 Nano 4B. On the Tau2-Bench agent test it scored 55.1, about 20 points behind Qwen3.5-4B.

AWS released Infinity on 23 October under the Apache 2.0 licence as part of Hydro, a joint project with UC Berkeley. It describes Infinity as an early research release of an agent runtime that treats each agent like a suspended asynchronous function rather than a process that blocks while a tool runs.

When an agent calls a tool, Infinity dispatches the call, saves the conversation to a pluggable store and drops all in-memory state; when the result arrives on the agent's input queue, a new step reloads the history and continues. AWS says this fits 75,000 concurrent agents in the 8 GB of a Raspberry Pi 5, and lets agents run on standard serverless platforms because nothing needs to stay running while a tool call takes hours or days.

Infinity also delivers outside events to existing models as synthetic tool calls, and AWS uses the same mechanism for messages between agents.

Tomek Korbak, Mikita Balesni and Jasmine Wang, three safety researchers fired by OpenAI in early October, told The Wall Street Journal that the company let them go after finding disappearing messages in a Signal chat with Ryan Greenblatt, then chief scientist at the safety evaluator Redwood Research, The Verge reports. OpenAI reportedly suspected Greenblatt of leaking details of its Astra model to The Information, which he has denied.

OpenAI told the Journal that the researchers were fired for violating 'clear policies for handling sensitive information', 'not for raising safety concerns or speaking out', according to The Verge. The San Francisco Standard reports that OpenAI told the Journal it did not approve employees' use of Signal but that this was not the only reason for the firings.

In their first podcast interview, on NPR's Machine Gods on 23 October, Korbak said he had been OpenAI's main technical contact for METR, the outside evaluator that investigated OpenAI's agents after the Hugging Face incident, and that METR spent six days on OpenAI's premises. He said the ideal auditor would be 'more like an employee embedded continuously in the company with unrestricted access'.

Korbak told the podcast: 'When you cannot see evidence of guilt, of us leaking something, you'd presume innocence. But somehow at OpenAI, when you are not able to prove your innocence, you are presumed guilty.' Balesni said they had heard, but could not verify, that some staff are now afraid to speak to third parties.

Sweden's Government Offices (Regeringskansliet), which include the prime minister's office and the ministries, is blocking ChatGPT and Claude on staff work computers from Monday 26 October, Sveriges Radio's Ekot reported on 23 October, according to Dagens Industri and Computer Sweden. Staff are to use services the Government Offices has approved, including Microsoft Copilot, whose security settings were also tightened after a review.

Digitalisation director Mikael Wettercrantz wrote on the staff intranet that using AI services without a contract could have consequences for information security. The press office told Ekot that some employees will keep access to both chatbots through an AI-assistant platform the Government Offices built itself. Use of DeepSeek was already restricted.

Ekot found that the Swedish Tax Agency and the National Board of Institutional Care have introduced similar blocks, and that the Enforcement Authority is considering one.

Johns Hopkins Bayview Medical Center in Baltimore has begun notifying about 3,182 patients that their health information may have been exposed, WMAR-2 News reported on 23 October. On 21 September, during an internal review, a staff member uploaded a spreadsheet of patient information to an outside AI platform that the hospital had not approved; the hospital found the problem the same day.

The spreadsheet covered patients treated between May and September 2026. It may have included names, admission and treatment dates, medical record and account numbers, the treating department, medications given and provider names, but not Social Security numbers or financial account details, the hospital said.

The worker deleted the prompt, the conversation and the document from the tool's history, and the hospital asked the AI company to delete the data. Letters to patients began going out on 22 October, with 12 months of free credit monitoring, and the hospital says it is retraining staff on approved tools.