S3dat

Building What Should Exist.
We build technology, products and ideas for what comes next.

[S3dat @s3datlab]

Our read: We think model intake needs an artifact review as well as behavior tests. A strong benchmark score does not establish that the file came through a trusted pipeline.

Full story: https://t.co/TgllNDuqa5

[S3dat @s3datlab]

A clean test run can miss a changed model.

Microsoft described Lazarus, a research backdoor in altered model weights. Its detector caught 91.4% of edited checkpoints with 9.7% false positives; those rates need to be read together.

#Microsoft #AISecurity

[S3dat @s3datlab]

Key facts:
· Research disclosed 10 November 2026
· 91.4% detection in the reference-free evaluation
· 9.7% false-positive rate in that evaluation

[S3dat @s3datlab]

Our read: We would test this as a small routing component before expanding its authority. Faster scoring is useful when the application keeps a clear path for uncertain cases.

Full story: https://t.co/kj8ifYKS7p

[S3dat @s3datlab]

A model can choose without writing an answer.

Google opened the Gemini Decisions API for fixed-choice judgments. Gemini Decision 1 handles up to 1,000 questions per request; developers still define the options and what happens afterward.

#Gemini #AIAgents

[S3dat @s3datlab]

Key facts:
· November 11: Gemini Decisions API announced
· Up to 1,000 independent questions per request
· $0.10 per million input tokens; zero output tokens
· 256K-token maximum context

[S3dat @s3datlab]

Our read: We favor testing hosted access before budgeting a local deployment. A working API integration does not settle the rights for downloading and serving the checkpoint.

Full story: https://t.co/HSUa5DUhbf

[S3dat @s3datlab]

The free endpoint is not the downloadable model.

Baidu opened ERNIE Flash Preview for hosted agent tasks. OpenRouter lists a free million-token endpoint; the existing base download has separate licensing terms.

#Baidu #ERNIE

[S3dat @s3datlab]

Key facts:
· November 11: Baidu's preview announcement
· 1 million tokens of context on OpenRouter
· Base license: $50 million commercial threshold

[S3dat @s3datlab]

Our read: We favor removing the dock requirement over chasing a headline frame rate. For handheld developers, readable interfaces and input latency belong in the same test.

Full story: https://t.co/c5IEOq8oa8

[S3dat @s3datlab]

The upscaler no longer needs a dock.

Microsoft released Auto SR for handheld play on battery, with 1080p game support. Its mode targeting up to 120 FPS arrives through Windows updates.

#Microsoft #AutoSR

[S3dat @s3datlab]

Key facts:
· November 10: handheld release leaves preview
· Ally X and X20 now support 1080p game input
· 120 FPS ceiling for the rolling-out Performance Mode

[S3dat @s3datlab]

Our read: We value the inspectable link between a security requirement and compiled code over proof volume. Human review of that requirement still decides what the result means.

Full story: https://t.co/xqunZ0GEqx

[S3dat @s3datlab]

A security proof still needs a human-approved promise.

DeepMind connected Signal security proofs to compiled code. Its four-week project used human-reviewed goals and machine checks; the result does not certify the whole messaging app.

#GoogleDeepMind #FormalVerification

[S3dat @s3datlab]

Key facts:
· Announced on 10 November 2026
· Project duration: 4 weeks, DeepMind says
· Signal described earlier verification on 2 October 2025

[S3dat @s3datlab]

Our read: We would evaluate saved worker state as a way to reduce idle capacity, with the storage bill and full request delay included in the comparison.

Full story: https://t.co/lGxwUsFHhz

[S3dat @s3datlab]

A saved setup can shorten the wait for a model.

NVIDIA announced Snapshot for restoring initialized GPU workers. Crusoe reports 31-second GLM-5.3 readiness on B300 hardware. The partner test does not establish fleet-wide savings.

#NVIDIA #Inference

[S3dat @s3datlab]

Key facts:
· Announced as a DSX component on 10 November
· Crusoe: 31 seconds from pod creation to ready
· Saturn: 6–34 seconds to first response in its tests

[S3dat @s3datlab]

Our read: We would use the fixed preview window for a repeatable evaluation, keeping production traffic on a separately supported model and billing route.

Full story: https://t.co/gy42FqbSI5

[S3dat @s3datlab]

A new model arrives with an expiry date.

Alibaba opened Qwen4-Max in preview on November 11. The version retires after November 25. Token Plan access is for interactive coding and agent tools, excluding application backends.

#Qwen #CodingAgents

[S3dat @s3datlab]

Key facts:
· Preview opened on 11 November 2026
· This version retires after 25 November
· Personal Edition excludes application backends