Iganädalane ülevaade
AI uudised
kolmapäev, 22. aprill 2026
Models
Meta Releases Llama 4 With Native Multimodal Reasoning
Llama 4 arrives with built-in vision and audio understanding, letting developers build multimodal apps on an open-weight model for the first time at this capability level.
Allikas: The Verge →GPT-5 Sets New Benchmarks Across Reasoning Tasks
OpenAI's latest model surpasses human expert performance on a suite of graduate-level reasoning benchmarks, with particular gains in multi-step logic and code generation.
Allikas: The Verge →Google DeepMind Releases Gemini Ultra 2 with 2M Token Context
The updated Gemini Ultra model doubles its context window, enabling analysis of entire codebases, lengthy legal documents, or full financial reports in a single prompt.
Allikas: Google Blog →Research
AI Hallucination Rates Drop 60% With New Retrieval Techniques
A joint paper from MIT and DeepMind shows that combining structured retrieval-augmented generation with lightweight verification agents cuts factual errors dramatically across enterprise use cases.
Allikas: MIT Technology Review →New Study: AI Assistants Cut Knowledge Worker Hours by 30%
A Stanford study tracking 2,000 professionals over six months found that regular AI assistant users reclaimed an average of 12 hours per week on research, drafting, and summarisation.
Allikas: Stanford HAI →Industry
LinkedIn Launches AI Career Coach Built on GPT-5
The feature, rolling out to Premium subscribers globally, offers personalised job search strategy, CV rewriting, and interview prep — positioning LinkedIn as an AI-first career platform.
Allikas: TechCrunch →Salesforce Embeds AI Agents Across Its Entire CRM Platform
Agentforce 3.0 brings autonomous AI agents into every Salesforce product — from lead scoring to contract drafting — with no-code configuration for non-technical teams.
Allikas: TechCrunch →Mistral Launches On-Prem Enterprise AI Suite
Mistral's new enterprise package lets companies deploy frontier-class models entirely on their own infrastructure, addressing data sovereignty concerns blocking cloud AI adoption.
Allikas: VentureBeat →Tools
Zapier AI Actions Now Support Multi-Step Autonomous Workflows
Zapier's updated AI layer can now plan, execute, and self-correct multi-step automations using natural language goals — no manual trigger configuration required.
Allikas: VentureBeat →Clay.com Launches AI-Native CRM to Compete With HubSpot
The sales enrichment startup is betting that a data-first, AI-native CRM built around automatic lead enrichment and outbound sequencing will displace legacy incumbents.
Allikas: TechCrunch →Microsoft Copilot Gains Autonomous 'Deep Research' Mode
Copilot's new research mode can autonomously browse the web, compile citations, and produce structured reports — positioning it directly against Perplexity's Pro offering.
Allikas: The Verge →Policy
EU AI Act Enforcement Begins: What Businesses Need to Know
The EU's landmark AI regulation entered its first enforcement phase this month. High-risk AI systems now require mandatory registration and conformity assessments.
Allikas: Wired →OpenAI Launches $100M Enterprise Safety Fund
The fund will support third-party auditors, red-teaming organisations, and academic labs focused on evaluating frontier AI systems before they reach enterprise customers.
Allikas: Bloomberg →UK Government Publishes AI Adoption Roadmap for Public Sector
The roadmap outlines a three-year plan to integrate AI across NHS administration, HMRC tax processing, and national infrastructure planning — with £800M in committed funding.
Allikas: GOV.UK →Funding
Anthropic Raises $3B to Accelerate Claude Development
The latest funding round values Anthropic at $45B and will go toward expanding compute infrastructure and advancing interpretability research.
Allikas: Bloomberg →