Benchmarks Topic - AI Updates

Abstract editorial illustration of an open geometric lattice and flowing translucent forms, representing a model preview and future access to weights

Mistral's trillion-parameter preview reports cyber and coding results while training continues. Its weights and private deployment options are planned for later in October.

An automated AI research system iterating through safety experiments under human supervision

View Post

AGI & Alignment

Anthropic Tests an Automated Researcher for AI Alignment

An Anthropic-led experiment used AI to propose, test and refine alignment interventions across ten benchmarks, while exposing the limits of automated research.

A sealed government evaluation room with abstract AI test patterns visible only through a narrow window

View Post

Regulation & Government

A Secret US AI Cyber Benchmark Raises a Public Accountability Problem

The White House plans voluntary pre-release cyber tests for advanced AI models, but classified criteria could favor established labs.

Scientific laboratory evaluating a glowing AI model core

View Post

AI Safety Is Becoming a Measurement Science

July 24, 2026

AI safety is moving toward a discipline built around repeatable measurements rather than broad claims about whether a model is safe or unsafe. Model evaluations can test capabilities, failure modes…

Powerful agentic AI model completing a multi-step terminal workflow inside a safe environment

View Post

Foundation Models

GPT-5.5 Raises Agent Capability—and the Price of Using It

GPT-5.5 posts strong terminal, coding and long-context results, but enterprises must test whether fewer retries justify higher token prices.

Researcher examining incomplete AI safety measurements beside rapidly advancing model capability

View Post

AI Safety & Security

Stanford Warns AI Safety Benchmarks Are Falling Behind

Stanford's 2026 AI Index finds inconsistent safety reporting, rising incidents and declining organisational confidence in incident response.

Office workers comparing AI assistance across documents, charts, summaries, and multiple languages

View Post

Rankings & Benchmarks

Samsung’s TRUEBench Tests AI on Real Office Work

TRUEBench evaluates language models across multilingual, context-rich workplace tasks instead of relying on academic question-and-answer tests.

Two researchers comparing benchmark results on ordinary monitors in a daylight office with green-blue undertones.

View Post

Open Source AI

DeepSeek V3-0324 Led Open-Weight Non-Reasoning Models

DeepSeek’s V3-0324 update became the first open-weight release to lead a prominent non-reasoning model index, pairing high capability with permissive access.

Advanced language model evaluated across reasoning coding and vision benchmarks

View Post

Foundation Models

Claude 3.5 Sonnet Raises the Bar for Speed, Cost and Capability

Anthropic’s mid-tier Claude 3.5 Sonnet outperformed larger models across key evaluations while doubling speed and strengthening its enterprise appeal.

A compact model represented by a small precise mechanism outperforming larger structures

View Post

Foundation Models

Phi-2 Shows How Better Data Can Let a Small Model Compete

Microsoft’s 2.7-billion-parameter Phi-2 used curated and synthetic training data to challenge models many times larger.

Follow AI Updates

Benchmarks

Mistral Large 4 Preview Separates Benchmark Claims From Its Planned Open-Weight Release

Anthropic Tests an Automated Researcher for AI Alignment

A Secret US AI Cyber Benchmark Raises a Public Accountability Problem

AI Safety Is Becoming a Measurement Science

GPT-5.5 Raises Agent Capability—and the Price of Using It

Stanford Warns AI Safety Benchmarks Are Falling Behind

Samsung’s TRUEBench Tests AI on Real Office Work

DeepSeek V3-0324 Led Open-Weight Non-Reasoning Models

Claude 3.5 Sonnet Raises the Bar for Speed, Cost and Capability

Phi-2 Shows How Better Data Can Let a Small Model Compete

inDrive’s Kazakhstan Robotaxi Plan Begins With a Feasibility Test

Aged Intelligence Society Makes AI Learning a Shared Exercise

Microsoft’s New Diagnostic Results Face a Real-World Test

Uber Expands ChatGPT Work, With Outcomes Still to Be Measured

Spotify Bets on AI Agents as a Distribution Channel

AWS Finds PHP Attack Attempts Before Public Disclosure

Meta Links Its Non-Voting Share Plan to Founder Leadership

Stay Ahead of AI

Follow AI Updates