AI engineering, November 8, 2026

Top News · 155 HN points

Agents do not automatically use what they remember. Anthropic nearly doubled Sonnet 4.6’s CL-Bench score by explicitly prompting it to read its memory and test its theories. Its broader memory design separates logging experience from idle-time “dreaming” that reviews logs, corrects mistakes, and prepares notes for future sessions.

Here is what makes this useful:

• A wiki built from about 350,000 internal documents rose from 13% to 48% of graded criteria across 25 questions as dreaming increased from two to 128 sessions.
• Search plus the growing wiki scored 61%, versus 53% for search alone, and saved about 20,000 tokens per question.
• Merger questions needing cross-document synthesis scored 87% with the wiki, versus 73% with search alone.

One catch: Memory instructions do not transfer cleanly across models. Telling Opus 5 and Opus 4.7 to record only failures reduced much of their memory benefit.

Sources: anthropic.com

Top News

SageMaker Inference can now place multiple models on hardware-isolated slices of one GPU. In AWS’s Llama-3.1-8B-Instruct test, an H100 partition cost 18% to 34% less per million output tokens than a whole A100 above a workload-dependent crossover.

Here’s what changed:

• The tested H100 mig-3g.40gb partition delivered 53% to 89% more output tokens per second at each traffic scenario’s greatest cost benefit.
• Each slice gets dedicated compute, memory, cache, and memory bandwidth.
• Partitioning is available through SageMaker inference endpoints in US East (N. Virginia and Ohio), US West (Oregon), Asia Pacific (Singapore and Tokyo), and Europe (Frankfurt and Ireland).

One catch: You still pay for the whole instance, and the A100 was cheaper below the crossover. Support is limited to real-time inference component endpoints. Changing the partition layout requires a new endpoint.

Try it: Set AcceleratorPartitionLayouts on the endpoint production variant and AcceleratorPartitionType on each inference component.

Sources: aws.amazon.com

Top News · Reported · semianalysis.com

Galaxy is Envision’s datacenter campus in Ulanqab, Inner Mongolia. SemiAnalysis reports up to 800MW of planned IT capacity for two internet giants, comparable with the roughly 820MW of critical IT capacity at Crusoe’s Abilene campus.

Here’s what is taking shape:

• The first 120MW phase was delivered in August 2026, with the full campus expected by the end of 2027.
• Envision’s wider Inner Mongolia datacenter plans total 3GW.
• Chinese datacenter rents have roughly halved in five years to about $40/kW/month excluding power, according to SemiAnalysis.
• The report says Galaxy’s factory-built substation took less than three months to connect, versus about six months for a conventional indoor substation.

One catch: Galaxy’s buildout is planned for up to 800MW. Abilene is also set to grow, with a planned expansion to 1.5GW.

Sources: newsletter.semianalysis.com

1. HyperFlow H3 Flash generates video and synchronized stereo audio in 6 steps · 59 HN points
2. Memorilla compresses document collections into 16 query-conditioned memory tokens for frozen Qwen3-8B
3. XMimic trains Unitree G1 motion-tracking policies in under a minute on four H100s · 60 Stars
4. OpenAI introduces BioBasicsBench, a 100-task benchmark testing AI agents on wet-lab science
5. LangChain's Managed Deep Agents adds browser use for websites with no API
6. Arete-27B cuts answers labeled propaganda from 22% to 0.9% in Nous Research's evaluation · 989 Downloads