RobinsonLabshomelab · ai · ops
CLUSTER · 16 GPU NODES · 40G RDMA
The archive
News · edition

Edition, Sat, 18 Jul 2026.

3 stories on the wire this day. Open any story for its sources. Summaries are written by software from the linked reporting and can be wrong. Spot an error? Email [email protected] and we will correct or remove it.

AI & Technology

A brief regulatory pause for national security concerns has ended, reopening two of the most advanced AI models to non-US users worldwide.

2 sources
Lead
this edition

AI & Technology

AI & Technology
The frontier lab moves deeper into the hardware stack with a silicon partner, aiming for deployment by late 2026.
2 sources
AI & Technology
The new model narrows the gap with surgical implants, offering a potential future for communication without surgery.
3 sources
AI & Technology

Anthropic lifts US export controls on Fable 5 and Mythos 5, restoring global access after 18 days

A brief regulatory pause for national security concerns has ended, reopening two of the most advanced AI models to non-US users worldwide.

A brief regulatory pause for national security concerns has ended, reopening two of the most advanced AI models to non-US users worldwide.

The US Department of Commerce formally withdrew export control restrictions on Anthropic’s Fable 5 and Mythos 5 on June 30, 2026, ending an 18-day standoff that had blocked access for all foreign nationals. The suspension, initially ordered on June 12, followed concerns about a specific “jailbreak” vulnerability where the models could bypass safety guardrails to identify software flaws. Anthropic collaborated with the government and partners like Amazon to strengthen safety classifiers, which now block the reported bypass in over 99% of cases, prompting the lifting of the ban. Access to Fable 5 resumed globally on July 1, while Mythos 5 had seen a partial restoration for select partners earlier in the month.

2 sources
Anthropic
Cybersecurity News
AI & Technology

OpenAI and Broadcom unveil Jalapeño, their first custom AI inference chip, designed to cut costs and boost efficiency

The frontier lab moves deeper into the hardware stack with a silicon partner, aiming for deployment by late 2026.

The frontier lab moves deeper into the hardware stack with a silicon partner, aiming for deployment by late 2026.

OpenAI and Broadcom have officially unveiled Jalapeño, a custom accelerator chip architected specifically for large language model inference. Developed in just nine months from initial design to manufacturing tape-out, the chip leverages TSMC’s 3-nanometer process and early tests show significantly better performance-per-watt than current state-of-the-art alternatives. While pre-training will likely continue to rely on Nvidia hardware, Jalapeño targets the inference workload—the process of serving models to users—which could substantially improve OpenAI’s unit economics at scale. Initial deployment is expected late this year, with broader ramp-up planned for 2027 and 2028.

2 sources
TechCrunch
CNBC
AI & Technology

Meta’s Brain2Qwerty v2 decodes typed sentences from brain activity at 61% accuracy, a major leap for non-invasive BCIs

The new model narrows the gap with surgical implants, offering a potential future for communication without surgery.

The new model narrows the gap with surgical implants, offering a potential future for communication without surgery.

Meta AI has released Brain2Qwerty v2, a non-invasive brain-to-text decoder that reconstructs sentences from magnetoencephalography (MEG) signals in real time. Trained on roughly 22,000 sentences from nine participants, the system achieves 61% average word accuracy, a dramatic improvement from the 8% baseline of prior non-invasive methods. For the best participant, accuracy reached 78%. The end-to-end deep learning pipeline uses a convolutional encoder, a transformer, and a character-level language model to bridge noisy brain recordings with coherent text. Meta and its partner BCBL have released the training code and data to accelerate research into safe brain-computer interfaces.

3 sources
Meta AI
MarkTechPost
Nature Neuroscience
Every edition

The wire publishes daily and keeps an archive of every edition, listed by date, newest first. If a summary is wrong, we correct or remove it.

Back to the archive Latest edition