Skip to main content

AI

AI model releases, infrastructure, security and policy, reported from the primary document: the release notes, the filing, the paper or the official post. Each article says what changed, why it matters and what to check next.

A collection of 14 posts

Databricks says AI Search can now index streaming tables, materialized views and managed Iceberg v3+ tables directly
AI

Databricks says AI Search can now index streaming tables, materialized views and managed Iceberg v3+ tables directly

Which table types Databricks AI Search can index now, what Databricks says replaces the old Delta copy and sync pipeline, and the materialized view setting.

· ZIPMEX

Dharma-AI's VocabX tokenizer: what the 77% inference cost cut measured
AI

Dharma-AI's VocabX tokenizer: what the 77% inference cost cut measured

Dharma-AI says its VocabX tokenizer cut a Qwen3-4B essay grader's measured inference cost 77.0%. What was tested, why it worked, and its limits.

· ZIPMEX

IBM Granite Speech 5.0: which model to use for transcription, translation and meetings
AI

IBM Granite Speech 5.0: which model to use for transcription, translation and meetings

IBM's Oct. 21, 2026 Granite Speech update: 5.0-2b for accuracy and translation, 470m TurboCTC for speed, 4.1-2b-plus for meetings.

· ZIPMEX

IBM's MIA plugin for vLLM: steering model internals with CUDA graphs still on
AI

IBM's MIA plugin for vLLM: steering model internals with CUDA graphs still on

IBM's open-source MIA plugin reads and steers vLLM model internals without eager mode. What it does, IBM's speed claims and what it requires.

· ZIPMEX

NVIDIA Confidential Bare Metal: what it is, what runs it, what's still unknown
AI

NVIDIA Confidential Bare Metal: what it is, what runs it, what's still unknown

NVIDIA's Confidential Bare Metal (Oct 21, 2026) lets model builders run their own stack on rented GPUs while keeping models and data protected.

· ZIPMEX

NVIDIA plans 2.5x substrate capacity in three years, with partner targets running to 2028
AI

NVIDIA plans 2.5x substrate capacity in three years, with partner targets running to 2028

NVIDIA's Oct. 21, 2026 plan: 2.5x substrate capacity in three years, doubled testing and partner dates from Feb 2027 to end-2028.

· ZIPMEX

Databricks MemCurrent: agent pass rate rises to 68.4% from 22.3% in its own test
AI

Databricks MemCurrent: agent pass rate rises to 68.4% from 22.3% in its own test

Databricks says MemCurrent memory raised AI agent pass rate from 22.3% to 68.4% on its internal benchmark. What was tested, what drove it, and the caveats.

· ZIPMEX

Apple's SelfCompact: when AI agents should summarize their own context
AI

Apple's SelfCompact: when AI agents should summarize their own context

Apple and Johns Hopkins researchers say an AI agent compacts its context better with a short rubric that says when to summarize and when to wait.

· ZIPMEX

Mix or tune? Apple researchers say adding English beat hyperparameter tuning for Arabic pre-training
AI

Mix or tune? Apple researchers say adding English beat hyperparameter tuning for Arabic pre-training

Apple-linked researchers ran about 1,000 Arabic pre-training runs. Mixing in English beat hyperparameter tuning, and the lead grew with model size.

· ZIPMEX

Microsoft Maia 300 explained: 2027 deployments and a multi-gigawatt plan
AI

Microsoft Maia 300 explained: 2027 deployments and a multi-gigawatt plan

Maia 300 is Microsoft's next AI inference chip, planned for production in 2027. Specs, rack design and how to read Microsoft's efficiency claims.

· ZIPMEX

NVIDIA and Cisco USFP: 400Gb/s-per-lane optics, 4x bandwidth, about 1.3x efficiency
AI

NVIDIA and Cisco USFP: 400Gb/s-per-lane optics, 4x bandwidth, about 1.3x efficiency

NVIDIA and Cisco's planned USFP optics run 400Gb/s per lane: 4x the bandwidth of 200G modules, ~1.3x the efficiency, no ship date yet.

· ZIPMEX

Cloudflare Managed Defense beta: how its AI alert reviews work
AI

Cloudflare Managed Defense beta: how its AI alert reviews work

Cloudflare's Managed Defense beta collects evidence with fixed code, then runs four AI agents and checks their citations. A human analyst still decides.

· ZIPMEX

Databricks Apps on-behalf-of-user authorization is GA: how it works and how to set it up
AI

Databricks Apps on-behalf-of-user authorization is GA: how it works and how to set it up

Databricks Apps can now act as the signed-in user, with Unity Catalog permissions applied. How OBO works, which scopes to add and what admins control.

· ZIPMEX

AstaBrief 8B: Ai2's open model for cited research reports
AI

AstaBrief 8B: Ai2's open model for cited research reports

Ai2's open AstaBrief 8B writes cited research reports 3.5x faster than Asta's Claude-powered mode. Benchmarks, training data, limits and how to run it.

· ZIPMEX