danielkrause/BSN-Chatsystem
1
0
Fork 0
Code Issues 14 Pull Requests Actions Packages Projects Releases Wiki Activity
Files
master
BSN-Chatsystem/morning_briefing_2026-06-07.md
T

10 KiB
Raw Permalink Blame History

Morning Briefing: KI/LLM-News - 7. Juni 2026 (Sonntag)

1. Claude Opus 4.8 veroffentlicht - neue Bestwerte bei Coding und Agentic Tasks

Quelle: https://www.anthropic.com/news/claude-opus-4-8 Datum: 28. Mai 2026

Teaser (DE): Anthropic hat Claude Opus 4.8 veroffentlicht. Das neue Modell verbessert Coding, agentische Fahigkeiten und Reasoning gegenuber Opus 4.7 deutlich, bei gleichem Preis. Neu: Nutzer konnen den Effort-Level steuern, Claude Code erhalt Dynamic Workflows fur grosse Projekte, und der Fast Mode ist jetzt 3x gunstiger. Auf dem Super-Agent-Benchmark ist Opus 4.8 das einzige Modell, das alle Cases end-to-end abschliesst und ubertrifft GPT-5.5 bei Kostenparitat.

Volltext (Auszug):

We are upgrading Claude Opus to a new version: Claude Opus 4.8. It builds on Opus 4.7 with improvements across benchmarks, and is a more effective collaborator. It is available today for the same price. Opus 4.8 launches alongside several new features. Users on claude.ai now have control over the amount of effort Claude puts into a task. Claude Code has a new dynamic workflows feature that allows it to tackle very large-scale problems. And fast mode for Opus 4.8, where the model can work at 2.5x the speed, is now three times cheaper than it was for previous models.

Claude Opus 4.8 has noticeably better judgment. In Claude Code, it asks the right questions, catches its own mistakes, pushes back when a plan is not sound, and builds up confidence around complex, multi-service explorations before making big changes. - Tom Pritchard, Staff Engineer

On our Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost. - Kay Zhu, Co-Founder and CTO

On CursorBench, Claude Opus 4.8 exceeds prior Opus models across every effort level. Tool calling is meaningfully more efficient, using fewer steps for the same intelligence.

2. Anthropic sammelt $65 Mrd. in Series H - Bewertung: $965 Mrd.

Quelle: https://www.anthropic.com/news/series-h Datum: 28. Mai 2026

Teaser (DE): Anthropic hat eine $65 Milliarden Series-H-Finanzierungsrunde abgeschlossen, angefuhrt von Altimeter, Dragoneer, Greenoaks und Sequoia. Post-Money-Bewertung: $965 Milliarden. Der Run-Rate-Umsatz hat $47 Mrd. uberschritten. Strategische Infrastrukturpartner wie Micron, Samsung und SK hynix sind mit an Bord, plus $15 Mrd. von Hyperscalern (inkl. $5 Mrd. von Amazon).

Volltext (Auszug):

Anthropic has raised $65 billion in Series H funding led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money. Global enterprises across industries are deploying Claude in their core operations. Since our Series G in February, adoption has continued to grow across global enterprise customers, and our run-rate revenue crossed $47 billion earlier this month.

This latest funding is expected to advance our safety and interpretability research, expand compute to meet growing demand for Claude, and scale the products and partnerships our customers rely on.

Claude is increasingly indispensable to our growing global community of customers, and we work tirelessly to make tools like Claude Code and Cowork more helpful, more powerful, and more adaptable to their needs, said Krishna Rao, CFO of Anthropic.

The round includes $15 billion of previously committed investments from hyperscalers, including $5 billion from Amazon. Joining them are strategic infrastructure partners: Micron, Samsung, and SK hynix, whose technologies play a critical role in the world supply of memory, storage, and compute.

3. Anthropic reicht vertraulichen S-1-Antrag bei der SEC ein (IPO-Vorbereitung)

Quelle: https://www.anthropic.com/news/confidential-draft-s1-sec Datum: 1. Juni 2026

Teaser (DE): Anthropic hat bei der US-Borsenaufsicht SEC vertraulich einen S-1-Registrierungsantrag fur einen Borsengang eingereicht. Anzahl und Preis der Aktien stehen noch nicht fest. Der IPO hangt von Marktbedingungen ab - ein klares Signal, dass der KI-Konzern den nachsten grossen Schritt Richtung offentliche Markte plant.

Volltext:

Today, Anthropic, PBC confidentially submitted a draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission for a proposed initial public offering of our common stock. This gives us the option to go public after the SEC completes its review. The proposed initial public offering will depend on market conditions and other factors. The number of shares to be offered and the price have not yet been set.

4. Project Glasswing wird massiv ausgebaut - 150 neue Organisationen

Quelle: https://www.anthropic.com/news/expanding-project-glasswing Datum: 2. Juni 2026

Teaser (DE): Project Glasswing, Anthropics Programm zur Sicherheitsanalyse kritischer Software mit Claude Mythos Preview, wird auf rund 150 weitere Organisationen in uber 15 Landern ausgeweitet. Neue Branchen: Strom, Wasser, Gesundheit, Telekommunikation und Hardware. Die ersten 50 Partner fanden bereits uber 10.000 Schwachstellen mit hohem oder kritischem Schweregrad. Anthropic warnt: Innerhalb von 6-12 Monaten werden viele andere KI-Firmen Modelle mit ahnlichen Cyber-Fahigkeiten haben, moglicherweise ohne Sicherheitsvorkehrungen.

Volltext (Auszug):

Project Glasswing is our collaborative effort to secure the most important software in the world. In early April, we announced that roughly 50 initial partners had access to Claude Mythos Preview, and since then, they have been deploying the model to scan their codebases for vulnerabilities. These partners have so far found more than 10,000 high- or critical-severity security flaws.

We are now expanding Project Glasswing. Following several weeks of close collaboration with our partners, the security industry, open-source software maintainers, and the US government, we are extending the partnership to approximately 150 new organizations. Each one will need to meet our security requirements before they gain access.

The organizations in this new group are based in more than 15 countries. The group covers several industries that were not well represented in our initial cohort, such as power, water, healthcare, communications, and hardware. Many of the new partners are vendors that maintain codebases relied upon by lots of other organizations around the world, including governments.

Cheap, fast AI models with powerful cyber capabilities are around the corner. We want Project Glasswing to spur institutions toward operating norms that reflect this reality. Within 6 to 12 months, we expect that many other AI companies will have Mythos-class models, and they could release them without safeguards that prevent misuse.

To address the scale of this coming challenge, hundreds of thousands of organizations, researchers, and maintainers will likely need access to the most advanced cyber capabilities and tools available. We are working as quickly as we can to safely release Mythos-level capabilities in general access.

5. RL.cu: Komplettes LLM Reinforcement Learning in Pure CUDA

Quelle: https://github.com/KJLdefeated/RL.cu (via Hacker News) Datum: 7. Juni 2026 (HN: 1pt)

Teaser (DE): Ein beeindruckendes Open-Source-Projekt: RL.cu implementiert das komplette LLM-RL-Pipeline (GRPO-Training) in reinem CUDA, null PyTorch. Mit handgeschriebenen CUDA-Kernels (FlashAttention-2, RMSNorm, RoPE, SwiGLU, AdamW), einer vLLM-ahnlichen Inference-Engine mit Continuous Batching und Paged KV Cache. Erreicht 1,37x schnellere Wall-Clock-Time als TRL mit vLLM beim GRPO-Training von Qwen3-0.6B auf DeepMath-103K.

Volltext (Auszug):

RL.cu: LLM Reinforcement Learning in Pure CUDA - From Kernels to GRPO. Zero PyTorch. Full RL loop. 1.37x faster than TRL with vLLM.

A from-scratch implementation of the complete LLM RL pipeline: hand-written CUDA kernels, a vLLM-style inference engine with continuous batching, and GRPO training.

What We Built:

  • CUDA Kernels: FlashAttention-2, RMSNorm, RoPE, SwiGLU, Embedding, Sampler, AdamW, GRPO loss - all with forward AND backward passes
  • Model: Qwen3-0.6B full forward + backward pass, safetensors weight loading
  • KV Cache: Paged KV cache with block manager (same design as vLLM)
  • Inference Engine: Continuous batching, CUDA graph capture, two-phase scheduling
  • Training: SFT + GRPO with gradient checkpointing, mixed-precision AdamW

Weitere Kurzmeldungen

Erstellt am 7. Juni 2026 | Quellen: anthropic.com, huggingface.co, news.ycombinator.com (Algolia API), reddit.com/r/LocalLLaMA (Zugriff blockiert)

Reference in New Issue View Git Blame Copy Permalink