Live News Feed
Unsere Live-News-Pipeline überwacht Entwicklungen in der AI-Branche in Echtzeit. Breaking-Meldungen erscheinen hier, sobald sie verifiziert sind.
7. Juni 2026
Lindy leitet 100 % des Traffics auf DeepSeek v4 um und verlässt Anthropic
Lindy CEO Flo Crivello has switched 100% of Lindy's production traffic to DeepSeek v4, churning from Anthropic models — saving millions of dollars while reporting performance improvements on core use cases.
LlamaIndex-CEO: Framework-Zeitalter ist vorbei — Schwenk zu verwalteter Infrastruktur
LlamaIndex CEO Jerry Liu declares the AI framework era over: the orchestration patterns LlamaIndex once wrapped have migrated into Claude Code, OpenClaw, and MCP tools. LlamaIndex is pivoting fully to LlamaParse (extraction accuracy) and LlamaCloud (managed service).
NVIDIA RTX Spark: 128 GB Unified Memory für lokale KI-Modelle
NVIDIA announces RTX Spark, a desktop/laptop mini-superchip with Blackwell GPU cores (~6,100), up to 20-core CPU, ~1 PFLOP FP4, and up to 128GB unified memory — enough to run frontier open models locally. Ships fall 2026.
ByteDance veröffentlicht Bernini als Open Source: Einheitlicher KI-Videoeditor
ByteDance has open-sourced Bernini, a unified video editing model comparable to Gemini Omni for video: edit by text, image, or video reference — add/remove characters, change background/season, and maintain multi-image character consistency. ComfyUI integration underway.
Town AI sammelt 55 Mio. USD Series A: 'Devin für alles andere'
Town AI has raised a $55M Series A led by Aaron Rampell at a16z, with Forerunner Ventures, First Round Capital, and Conviction. The startup's AI handles email, calendar, Slack, scheduling, and multi-step project tasks — acting only when the user approves.
Microsoft-KI entwirft 1.000-fach zuverlässigeren Quantenchip
Microsoft used its Discovery AI system to design a new quantum chip material stack reportedly 1,000x more reliable than prior qubit designs, halving its target timeline to a scalable quantum computer to 2029.
Google Magenta RealTime 2: 200 ms On-Device-Musikgenerierung
Google releases Magenta RealTime 2, the only open-weights model for real-time continuous music generation on-device: 200ms latency, steerable by text, audio, or MIDI, and runs on a MacBook (2.4B parameters, no GPU required).
OpenAI-Modell findet Gegenbeispiel zu 80 Jahre alter Erdős-Vermutung
Researchers working with an OpenAI model have found a counterexample to an 80-year-old Erdős conjecture in mathematics — an example of AI-assisted mathematical discovery beyond benchmark performance.
Harvey + LangChain Labs: Rechtliche KI-Verifikation 1.000-fach günstiger
Harvey and LangChain Labs publish a legal AI verifier efficiency study: batch LLM-as-judge scoring reduces verification cost ~1,000x versus per-criterion calls. DeepSeek v4 Flash preserves 94-96% of Opus 4.7 verifier signal at 18x lower cost per criterion.
MiniMax M3: Open-Weights-Modell mit 1-Mio.-Token-Kontext führt Coding-Ranglisten an
MiniMax unveils M3: an open-source agentic-coding model with a 1M token context window, multimodal inputs, and SOTA performance on SWE-bench-Pro — beating GPT-5.5. API available at ~$0.20/M tokens; weights release in 1-2 weeks.
ParseBench auf der CVPR 2026: Erste KI-Agenten-Dokumentenbenchmark
LlamaIndex has presented ParseBench at CVPR 2026 — the first document-understanding benchmark designed for AI agents, covering 2,000+ human-verified pages, 167K+ test rules, and 5 evaluation dimensions. Fully open source.
LangChain startet LangSmith-Plattform-Offensive: Fleet, Engine, Gateway
LangChain shipped a coordinated product blitz: LangSmith Fleet (no-code natural-language agent builder), LangSmith Engine (automated trace-to-issue-to-fix-to-eval loop), LangSmith Sandboxes GA (stateful CLI-accessible compute), and LLM Gateway (traceable governance inside LangSmith).
Ideogram 4.0 Open Weights veröffentlicht: Bestes offenes Bildmodell
Ideogram AI has released open weights for Ideogram 4.0, claiming the #1 position among open image models. Weights are downloadable, fine-tunable on custom data, and available at huggingface.co/ideogram-ai/ideogram-4-nf4.
Claude Code Dynamic Workflows: KI schreibt eigene Orchestrierungssysteme
Claude Code Dynamic Workflows, live since May 28, allow Claude to author its own JavaScript orchestration harness with subagents, parallel execution, per-agent model selection, and agent isolation — giving AI structural control over its own execution architecture.
NVIDIA Cosmos 3 führt 7 Physical-AI-Ranglisten an
NVIDIA releases Cosmos 3, an open-weights physical AI foundation model that ranks #1 on 7 leaderboards spanning world generation (PAI-Bench, Physics-IQ, R-Bench), robot action policy (RoboLab), and industrial vision (VANTAGE-Bench, TAR).
Perplexity startet Computer: Vollständig gehosteter Allzweck-KI-Agent
Perplexity has launched Perplexity Computer, a fully-hosted AI agent with threaded tasks, hundreds of pre-built connectors, custom skills, sub-agent delegation, and orchestrator model choice (Opus 4.6, GPT-5.4, or Sonnet 4.6).
Cognition bietet 10-Mio.-USD-KI-Produktivitätsgarantie für Devin
Cognition launches an 'AI Productivity Guarantee' for Devin: if the AI coding agent delivers less engineering value than the customer pays for, Cognition funds usage until it does — up to $10M. Enterprise private evals extend to 100 hours.
Gradio 6.16.0 schließt Path-Traversal-, OAuth- und SSRF-Lücken
Gradio 6.16.0 patches three security vulnerabilities: path traversal in gr.FileExplorer, an open-redirect bypass in OAuth, and SSRF in Image/Gallery/Audio post-processing. Upgrade immediately.
KI-CEOs unterzeichnen Biosicherheitsbrief an den US-Kongress
Sam Altman, Dario Amodei, and Demis Hassabis have co-signed a letter urging Congress to increase oversight of synthetic nucleic acid supply chains, citing AI's growing biological capabilities as a dual-use risk.
S&P DJI lehnt SpaceX-Schnellaufnahme in den S&P 500 ab
S&P Dow Jones Indices has formally rejected all proposals to fast-track SpaceX into the S&P 500, maintaining the 12-month IPO seasoning period, positive GAAP net income requirement, and 10% public float — SpaceX's earliest eligibility is now June 2027.
GitHub startet MCP Apps: Interaktive Benutzeroberfläche direkt im Chat
GitHub and Microsoft have unveiled MCP Apps, extending the Model Context Protocol to let server tools return rich interactive HTML components rendered in sandboxed iframes inside VS Code chat. Shopify, Excalidraw, and Figma are early adopters.
Stripe startet drei Agentic-Zahlungsprimitiven für KI-Agenten
Stripe unveils Shared Payment Tokens (scoped mandate credentials), Machine Payments Protocol (HTTP 402 pay-per-tool-call), and the OpenAI-co-built Agentic Commerce Protocol for deterministic agent checkout. Stripe Projects wraps all three.
Supabase sammelt 500 Mio. USD bei 10 Mrd. USD Bewertung ein
Supabase raises $500M at a $10B valuation, with an employee-friendly equity structure: 25% cashout option on vested shares and a 10-year exercise window compared to the standard 3 months.
Google veröffentlicht Gemma 4 12B: Encoder-freies multimodales Modell
Google releases Gemma 4 12B under Apache 2.0: unified encoder-free multimodal model with 256K context, native audio/image/video support, agentic reasoning — runs in 16GB VRAM with Day 0 Transformers, llama.cpp, and MLX support.
ChatGPT führt automatisches kontextbewusstes Gedächtnissystem ein
OpenAI launches a new automatic memory system for ChatGPT with 2x more capacity, rolling out to US Plus and Pro users. The system tracks context dynamically across conversations and adapts over time to temporal changes.
Claude Mythos erreicht 3-Stunden-Horizont für autonome Aufgaben
Claude Mythos achieved a 3-hour 6-minute autonomous task horizon on METR's benchmark in late May — hitting the median superforecaster prediction for end of 2026, months ahead of schedule.
Anthropic stellt vertraulichen Börsengangsantrag bei ca. 965 Mrd. USD Bewertung
Anthropic has filed for an IPO confidentially at approximately $965B valuation. President Daniela Amodei cites the high cost of developing AI models as the primary driver for going to public markets.
Anthropic: 80 % des eingepflegten Codes jetzt von Claude geschrieben
Anthropic publishes internal data: >80% of merged code is Claude-authored, engineers ship 8x more code, open-ended task success jumped from 26% to 76% in 6 months, and a training optimization reached 52x speedup.
Microsoft startet MAI: Sieben selbst entwickelte KI-Modelle
Microsoft ships seven MAI models trained from scratch on commercially licensed data, headlined by MAI-Thinking-1 (35B, 97% AIME, 53% SWE-Bench Pro) — a strategic bid for frontier model independence from OpenAI.
NVIDIA veröffentlicht Nemotron 3 Ultra: 550-Milliarden-Parameter Open-Source-Modell
NVIDIA releases Nemotron 3 Ultra — a 550B-parameter MoE model with 1M token context, 5x faster inference, and full open weights under OpenMDW-1.1 on Hugging Face.