AI Frontier Digest

New Benchmarks

• 🔥 HorizonMath: HorizonMath is a new benchmark featuring over 100 predominantly unsolved problems in computational and applied...

Pushing beyond fine-tuning dominance:

• Qianfan-OCR delivers a unified end-to-end model for document intelligence
• Fine-tuning seems cheaper but...

Google's agentic AI trend boosts tech worker productivity:

• Sashiko launch: Google engineers' tool for agentic AI code review on Linux Kernel,...

Microsoft has hired the full team from Sequoia-backed AI collaboration startup Cove, which is shutting down with service ending April 1 and customer data set for deletion.

Trend alert: Policy exclusions and technical failures expose urgent AI agent reliability gaps in deployment.

• DOD exclusion: Anthropic labeled...

Rox AI sprints to a $1.2 billion valuation, cementing unicorn status in AI-driven sales automation since its 2024 founding. This fuels the accelerating AI race for agent-powered sales tools.

Arena, formerly LM Arena, is the de facto public leaderboard for frontier LLMs, influencing funding, launches, and PR cycles.

• Billed as the...

Manus AI agent's rapid timeline:

• Desktop launch announced amid top AI stories.
• Meta acquisition escalates; China restricts Manus executives from...

Agentic AI targets clinical trial chaos:

• Zurich startup Rivia raises €13M after €3M to unify fragmented data.
• Builds AI agents to actively manage...

Saudi Aramco's $500M VC arm Wa’ed Ventures makes strategic investment in Resemble AI to expand deepfake detection capabilities across the Middle East. Vital boost for AI ethics and security from regional capital.

Together.ai's open-source Mamba-3 delivers a 4% edge over Transformers on language benchmarks while running 7x faster on long sequences via H100...

Emerging benchmarks highlight LLM struggles in advanced math, lagging humans and revealing eval flaws:

• MathVista: GPT-4V tops 12 models at 49.9% vs....

• Hands-on test of 5 LLMs: GPT-4.1, GPT-5.1, GPT-5.2, Claude 4.6, Gemini 3 on latency, function calling, instruction following, conversation ability,...

• PeruMedQA benchmark targets LLMs on Peruvian physician specialty exams, building a fine-tuned dataset to probe real gaps.
• Triage failures stem...

• Open-source framework for performance evaluation of widely used AI workloads in computer vision and large language models
• Focuses on power-aware...

Can AI reason about a war before its trajectory becomes historically obvious? New arXiv paper flags how retrospective analysis complicates benchmarking long-horizon strategic capabilities.

OpenAI's $110 billion funding round marks the largest AI investment in history, reshaping the AI investment landscape in spring 2026 and signaling major shifts ahead.

Emerging trend: New benchmarks target step-level quality for safe, heavy-duty AI agents.

• AgentProcessBench evaluates LLM tool-use in real-world...

NVIDIA and Mistral launch the Nemotron Coalition to deliver frontier-level open-source AI models, pitting Mistral Small 4 and DGX Cloud against...

A new cognitive framework proposes measuring progress toward AGI, sparking discussion with 49 points on Hacker News—a fresh angle beyond standard benchmarks for tech enthusiasts.