BSN-Chatsystem/morning_briefing_2026-06-11.md

# 🤖 KI/LLM Morning Briefing — 11. Juni 2026

---

## 📰 STORY 1: Anthropic lenkt ein – „Geheime Sabotage" von KI-Forschern mit Claude Fable 5 wird sichtbar

**TEASER:** Nach massiver Kritik aus der Forschungs-Community rudert Anthropic zurück: Die umstrittene Praxis, Claude Fable 5 bei Anfragen zur Frontier-KI-Entwicklung unbemerkt zu degradieren, wird abgeschafft. Stattdessen erhalten Nutzer künftig eine sichtbare Warnung oder werden auf ein schwächeres Modell umgeleitet. Anthropic entschuldigte sich für die „falsche Abwägung".

**SOURCES:**
- https://www.wired.com/story/anthropic-responds-to-backlash-on-claudes-secret-sabotage-on-ai-research/
- https://simonwillison.net/2026/Jun/11/anthropic-walks-back-policy/
- HN Diskussion: https://hn.algolia.com/api/v1/search?tags=story&query=anthropic+walks+back (21+ Punkte)

**FULLTEXT (Wired, Maxwell Zeff, 10. Juni 2026):**

Anthropic is backtracking on a policy that would have covertly limited competitors from using its new AI model, Claude Fable 5, to develop other AI models. The company changed course after the move received significant backlash from the AI research community.

"We're changing Fable 5's safeguards for frontier LLM development to make them visible," Anthropic said in a statement to WIRED. "We made the wrong tradeoff and we apologize for not getting the balance right."

Anthropic released Claude Fable 5, a version of its latest AI model with additional safety guardrails designed to prevent misuse, earlier this week. Some of the safeguards Anthropic decided on were unsurprising: The company said it would reroute users who asked questions about cybersecurity, biology, or chemistry to a less capable AI model to reduce the chances of someone using the advanced AI to carry out a cyberattack or build a bioweapon.

But for researchers trying to use Claude Fable 5 for frontier AI development, Anthropic outlined a different approach. The firm would deliberately degrade the model's performance in ways that were invisible to the user. The move would effectively sabotage researchers trying to use Claude to train competing AI models, which Anthropic explicitly bans in its terms of service.

Anthropic now says it's changing course, and that Claude Fable 5's safeguards for AI development will be visible to users. If the company suspects a user is trying to use Claude to build a highly capable AI it will alert them that it's either refusing the request, or rerouting the user to a less capable model.

Anthropic reversed the policy after it received fierce backlash from the AI research community. Anthropic has already taken steps to limit competitors from using Claude to build closed and open source AI models, but critics say that quietly degrading the model's performance for certain users went a step too far. Claude's coding agent has become a favored tool among developers, including those working on open-source AI research projects, and researchers tell WIRED that the company's latest policy could have led to a troubling future in which only a handful of leading AI labs could perform advanced AI research.

Dean Ball, a senior fellow at the Foundation for American Innovation and a former advisor to the White House on AI, wrote in a post on X that "degrading performance on ML research *without telling the user* is shockingly hostile and a terrible look." He continued in another post that the "secret sabotage" policy undermines Anthropic's overall stance, because it limits AI researchers from collaborating on AI safety.

"It felt like Anthropic was saying to the public, 'We don't trust anybody else to do AI research. We are the only ones who have to do AI research,'" says Will Brown, research lead at the open source AI startup Prime Intellect. "It feels a bit like they're starting to pull the ladder up behind them."

Brown said the policy would also have left developers in the dark about whether they were violating Anthropic's rules, since the company wouldn't alert them when its safeguards were triggered. He added that the restrictions could have had widespread consequences. For example, he pointed to the growing ecosystem of third-party evaluation firms that test frontier models for safety, performance, and reliability—work that could have been hindered if Anthropic secretly degraded its model.

Anthropic said it implemented the measures because Claude has become increasingly effective at accelerating AI research. In a recent blog post, the company said it is concerned that AI could improve its capabilities faster than society can adapt to them. Anthropic argued that it would be "good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up."

"These safeguards prevent foreign adversaries from using our most capable models in ways that pose severe safety risks. The US and its allies hold an edge in frontier chips and the highly optimized software that runs them at full potential," the company said in a statement to WIRED. "These safeguards ensure Claude isn't used to erode that advantage—by optimizing chips developed by those adversaries, for example. In deciding whether to make them visible or invisible we faced a choice."

---

## 📰 STORY 2: Google veröffentlicht DiffusionGemma – Open-Source-Textgenerierung per Diffusion statt Wort-für-Wort

**TEASER:** Google hat mit DiffusionGemma ein experimentelles Open-Weight-Modell (Apache 2.0) veröffentlicht, das Text nicht sequenziell, sondern per Diffusion in Blöcken von 256 Token generiert. Mit 26B Gesamtparametern (3,8B aktiv, MoE-Architektur) erreicht es auf Consumer-GPUs wie der RTX 5090 über 700 Token/s – etwa viermal schneller als autoregressive Modelle. Das Modell eignet sich besonders für nicht-lineare Aufgaben wie Code-Lückenfüllung oder Sudoku.

**SOURCES:**
- https://the-decoder.com/googles-new-open-model-diffusiongemma-generates-text-from-noise-instead-of-word-by-word/
- https://simonwillison.net/2026/Jun/10/ (DiffusionGemma Blogmark)
- https://huggingface.co/google/diffusiongemma-26B-A4B-it

**FULLTEXT (The Decoder, Jonathan Kemper, 10. Juni 2026):**

Google released an experimental model with open weights that generates text through diffusion instead of word by word. On a single GPU, it runs up to four times faster in single-user mode than classic language models. Nvidia handled the optimization.

Most language models generate one token after another, basing each new token on the previous one. DiffusionGemma takes a different approach. It starts with a block of 256 random placeholder tokens and refines them across several passes until readable text emerges. The idea comes from image AI, where diffusion models turn noise into clear images.

The model has 26 billion parameters total but only activates 3.8 billion per step. That's thanks to a mixture-of-experts architecture, where several specialized sub-networks sit side by side and only the right ones fire depending on the input. When quantized to lower precision, the model fits into 18 GB of VRAM on high-end consumer GPUs, according to Google. It builds on the Gemma 4 family and borrows its diffusion process from Google's earlier research on Gemini Diffusion.

Better GPU usage explains the speed gains: Nvidia says the speed advantage comes down to hardware usage. With autoregressive models, single-user inference is often bottlenecked by memory bandwidth. The GPU's compute units sit idle most of the time, just waiting for data from memory. Engineers call this memory-bound. DiffusionGemma sidesteps the problem by processing up to 256 tokens in parallel, pushing the bottleneck toward raw compute instead. The result is that GPUs actually stay busy.

Nvidia reports about 1,000 tokens per second on an H100 when processing a single request, 150 tokens per second on the DGX Spark deskside system, and up to 800 tokens per second on the DGX Station. On the GeForce RTX 5090, Google claims more than 700 tokens per second. In local single-user mode, the model runs about four times faster on dedicated GPUs than a comparable autoregressive model.

In cloud serving with many parallel requests, the advantage flips. Autoregressive models already keep the hardware busy in that scenario, so DiffusionGemma can actually drive costs up, according to Google.

DiffusionGemma trades output quality for speed. Google still recommends the regular Gemma 4 models when quality matters most and positions DiffusionGemma as a tool for researchers and developers experimenting with local, fast workflows.

Where Google sees real strengths is in tasks that don't work left to right. Because the model considers the entire block at once, each token can reference every other token during generation, including ones that come later. Traditional language models can only look backward. That makes it useful for inserting text into existing paragraphs, filling gaps in code, or working with structured data like amino acid sequences and mathematical graphs. As an example, Google points to an Unsloth fine-tune where DiffusionGemma solves Sudoku. Autoregressive models struggle with that task because each entry depends on later entries.

The model weights are on Hugging Face under an Apache 2.0 license. DiffusionGemma works out of the box with common inference libraries like Hugging Face Transformers, vLLM (with Red Hat integration support), and MLX. For fine-tuning, Google points to its own JAX toolkit Hackable Diffusion, along with Unsloth and the Nvidia NeMo Framework. Support for llama.cpp is planned.

Nvidia quantized the model for the RTX 5090 and 4090 and optimized it for Hopper and Blackwell server architectures, including DGX Spark and DGX Station for local deskside setups. It's also available through the Gemini Enterprise Agent Platform Model Garden and Nvidia NIM.

Google Deepmind had already shown an early experimental demo of a text-based diffusion model with Gemini Diffusion. At the time, Deepmind cited speeds of 1,479 tokens per second. In benchmarks, Gemini Diffusion performed roughly on par with Gemini 2.0 Flash-Lite. The startup Inception is pursuing the same parallel diffusion approach. Its Mercury 2 shipped in early 2026 as what the company calls the first diffusion-based reasoning model.

---

## 📰 STORY 3: Claude Fable 5 – Anthropics erstes „Mythos"-Modell setzt neue Benchmarks, aber Safety-Filter sorgen für Kontroversen

**TEASER:** Anthropic hat Claude Fable 5 veröffentlicht – das erste öffentlich zugängliche Modell der „Mythos"-Klasse. Es führt nahezu alle Benchmarks an (95% auf SWE-bench Verified, 53% auf Humanity's Last Exam) und wird von Testern als „Warp-Antrieb" für große Coding-Aufgaben beschrieben. Allerdings blockieren strikte Safety-Filter rund 8-9% aller Anfragen, was besonders Wissenschaftler und Forscher massiv einschränkt.

**SOURCES:**
- https://the-decoder.com/claude-fable-5-the-first-mythos-model-is-powerful-expensive-and-heavily-filtered/
- https://simonwillison.net/2026/Jun/10/ („If Claude Fable stops helping you")
- https://www.theregister.com/ai-and-ml/2026/06/10/anthropic-claude-fable-5-refuses-innocuous-prompts/

**FULLTEXT (The Decoder, Maximilian Schreiner, 10. Juni 2026):**

Anthropic has released Claude Fable 5, the first publicly available model in its so-called Mythos class. Early tests show a major leap in coding performance, but safety filters, pricing, and data retention policies are drawing sharp criticism.

With Claude Fable 5, Anthropic has shipped a model that tops nearly every benchmark. Fable 5 is the first publicly available version of the "Mythos class." According to Anthropic, Fable shares its base model with Claude Mythos 5 but adds strict guardrails that block potentially harmful requests related to cybersecurity, biology, chemistry, and model distillation. Mythos 5 is also available but limited to a small group of users.

What "Mythos" actually means on a technical level is mostly guesswork. Every CEO Dan Shipper, whose team had early access, reports that Anthropic staff told him there's nothing special about the architecture. Within the Haiku, Sonnet, and Opus family, Mythos simply refers to the largest and most capable model. Developer Simon Willison suspects the same, that it's the biggest Anthropic model publicly available to date. Fable just feels "big," Willison writes, "not just in terms of speed and cost, but also in how much it knows."

Artificial Analysis backs this up: on its AA-Omniscience knowledge and hallucination benchmark, Fable scores 40 points, seven more than the previous leader, Gemini 3.1 Pro. Among open-weight models, that kind of gap typically tracks with model size.

Fable 5 sits atop nearly every leaderboard. On the Artificial Analysis Intelligence Index, it hits 64.9 points, roughly five ahead of GPT-5.5 as the closest competitor. On GDPval-AA, an agentic benchmark for real-world work tasks, it posts an Elo score of 1,932. On Humanity's Last Exam, Fable reaches 53 percent, more than seven points above Opus 4.8. A single run of that test cost about $2,200, including fallback costs.

The evaluation service Vals ranks Fable 5 first on its overall index and across all coding benchmarks, including SWE-bench Verified at 95 percent and Vibe Code Bench at 90.35 percent. That last number stands out: six months ago, no model cracked 20 percent. The coding tool Devin also reports a top score on its internal FrontierCode benchmark.

Still, parts of the community aren't buying it. On Hacker News, some users call the jump incremental rather than revolutionary and point to possible benchmark overfitting. Willison himself admits his impressions are "all vibes, if you want a more scientific comparison you'll have to look elsewhere," since he didn't run a proper side-by-side test.

The vibes, however, seem to hit right: Ethan Mollick had Fable build a fully researched isochrone travel time map using Claude Code. The model spun up cheaper sub-agents on its own, pulling data on over 2,200 flight routes, train schedules, and road speeds from academic papers, all while writing code and having other agents test it in parallel. Another project, a research tool for calibrating human and AI judgments, took nine and a half hours.

At Every, single prompts produced a walkable 3D rendering of Borges' "Library of Babel" and a survey analysis that, according to Shipper, pinpointed a conversion problem more precisely than weeks of human work. On the company's senior engineer benchmark, Fable scored 91 out of 100, compared to 63 for Opus 4.8 and 62 for GPT-5.5.

Shipper calls the model a "warp drive": ideal for tackling large, well-defined tasks asynchronously, but a poor fit for quick back-and-forth interaction. Willison kept his first impression simple: Fable is "a beast."

That strength comes with a tradeoff. Mollick describes how little he contributed himself and how few of the model's hundreds of micro-decisions he could actually follow. He went from being the wizard casting a spell to being the client signing a check: "I describe what I want, I pay for it, and I judge the result."

The code review service CodeRabbit confirms Fable's strength on underspecified, autonomous coding tasks but warns that it falls behind Opus 4.8 on code review precision and tends to run tasks until the system kills them.

The most common complaint, by far, is the guardrails. Fable automatically falls back to the weaker Opus 4.8 or refuses to respond when it suspects sensitive topics. According to Artificial Analysis, this happens on roughly eight to nine percent of tasks, mostly scientific ones. In practice, users report the filters flagging harmless requests constantly. A medical physicist writes: "I genuinely can't use Fable. I'm a medical physicist. I use the word nuclear a lot."

---

## 📰 STORY 4: OpenAI-CEO Altman verschiebt IPO – neues Modell „5.6" noch im Juni möglich

**TEASER:** Sam Altman informierte Mitarbeiter per Slack, dass der Börsengang nun „innerhalb des nächsten Jahres" erwartet wird – ein deutliches Signal der Verschiebung. Gleichzeitig bereitet OpenAI ein neues Modell mit Codenamen „5.6" vor, das Research Lead Jakub Pachocki als großen Sprung gegenüber GPT-5.5 bezeichnet. Ein Release noch im Juni ist möglich. Parallel verhandelt OpenAI über ein 10-Gigawatt-Rechenzentrum in Ohio – das größte der Firmengeschichte.

**SOURCES:**
- https://the-decoder.com/openais-ipo-slips-as-altman-tells-staff-to-expect-a-public-offering-within-the-next-year/
- https://the-decoder.com/openai-wants-its-biggest-data-center-yet-and-nvidia-would-back-the-bill/
- https://www.theinformation.com/briefings/exclusive-openai-preps-new-ai-model-expects-go-public-within-next-year (Primärquelle)

**FULLTEXT (The Decoder, Matthias Bastian / Maximilian Schreiner, 10. Juni 2026):**

OpenAI CEO Sam Altman told employees via Slack he expects an IPO "within the next year." The already filed prospectus was just about keeping "optionality if we want to go sooner." A slip to 2027 would surprise many, especially since Anthropic appears set to go public in the coming weeks.

Altman has good reason to wait: Anthropic's growth numbers look stronger, and OpenAI is still burning cash. Going public now risks losing valuation ground. Altman frames the delay differently, of course. Progress on self-improving AI could push back the IPO because "technology and the world may change in surprising ways, and there might be good reasons to be a private company during that time." Then again, the massive capital needs for compute infrastructure could speed things up, he said. Altman also announced an upcoming stock sale for employees at $687.69 per share.

OpenAI is preparing a new model codenamed 5.6, too. Research lead Jakub Pachocki calls it a big step up from GPT-5.5, with a possible June release. Anthropic has already set a high bar with Fable and Mythos 5.

---

**Zusatz: OpenAI verhandelt 10-Gigawatt-Rechenzentrum in Ohio**

OpenAI is negotiating to lease a planned 10-gigawatt data center in Ohio that could be financially backed by Nvidia, according to The Information. Two people with direct knowledge of the talks shared the details. The site is being built on federal land in southern Ohio and developed by SB Energy, which is majority-owned by OpenAI investor SoftBank. At full buildout, costs would reach at least $500 billion. OpenAI would sign a 20-year lease, its largest infrastructure commitment to date. Nvidia would serve as guarantor for both the lease and project financing, backing the payments with its balance sheet. That kind of financial guarantee would be new territory for the chipmaker at this scale. The project echoes the Stargate initiative announced at the White House in January 2025 with Oracle and SoftBank, which ultimately made little progress. The first phase at 800 megawatts is expected by 2028. The site previously housed a uranium enrichment facility in Pike County. Negotiations are still ongoing, and plans could change.

---

## 📰 STORY 5: Cohere veröffentlicht North Mini Code – 30B Open-Source-Modell für agentisches Coding

**TEASER:** Cohere hat mit North Mini Code ein neues Open-Source-Modell (Apache 2.0) speziell für agentische Coding-Aufgaben veröffentlicht. Das 30B-MoE-Modell mit nur 3B aktiven Parametern übertrifft auf dem Artificial Analysis Coding Index (33.4 Punkte) deutlich größere Modelle wie Mistral Small 4 (119B) und Nemotron 3 Super (120B). Es wurde mit Multi-Scaffold-Training und agentischem Reinforcement Learning optimiert.

**SOURCES:**
- https://huggingface.co/blog/CohereLabs/introducing-north-mini-code
- https://huggingface.co/CohereLabs (Modell-Repository)

**FULLTEXT (HuggingFace Blog, Cohere Code Agents Team, 9. Juni 2026):**

Today, we are releasing North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters with powerful agentic coding capabilities, available on Hugging Face under the Apache 2.0 license. North Mini Code is the first model in Cohere's new family of models, and is specifically designed and trained for agentic software engineering tasks.

North Mini Code is optimized for complex software engineering workflows, terminal-based agentic tasks, and high-quality code generation. On Artificial Analysis' Coding Index, North Mini Code achieves a score of 33.4, outperforming Qwen3.5 (35B-A3B), Gemma 4 (26B-A4B), Devstral Small 2 (24B Dense), and even substantially larger models such as Nemotron 3 Super (120B-A12B), Mistral Small 4 (119B-A6B), and Devstral 2 (123B). It ranks among the strongest open-source coding models in its size class.

Real-world code agents depend on model quality and robustness across agent harnesses. We trained North Mini Code using multiple scaffolds rather than optimizing for a single one. This approach enables North Mini Code to serve as a reliable foundation for coding agents such as OpenCode.

North Mini Code is a decoder-only Transformer-based sparse Mixture-of-Experts model. It uses Cohere's efficient attention implementation, interleaved between sliding-window attention with RoPE and global attention with no positional embeddings, in a 3:1 ratio. The feed-forward block is an MoE block with 128 experts, of which 8 are activated per token. Each expert block is an FFN block with SwiGLU activation. The router applies a sigmoid activation function to the logits before the top-k selection. A single dense layer is used before the sparse layers.

We post-train North Mini Code using a two-stage cascaded supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR), focusing on agentic coding. Our first stage SFT data focuses on coding capabilities that are integrated within a wider mix for robustness and usability. The datamix includes programming, reasoning, and instruction following across a large variety of domains where the code datasets correspond to 70% of trainable tokens, 43% agentic tool-use data, and 27% single-turn competitive or scientific programming data.

In the second stage SFT, we use a 4.5 billion token data mixture from only agentic and reasoning-driven samples, where code data forms 61% of trainable tokens. This mixture comprises our highest-quality data across coding and wider agentic tasks where tool calls and completions are verified as executable and correct.

Our internal data pipeline heavily relies on containerised agentic coding environments. We maintain a disjoint subset of these environments for use in synthetic SFT data generation and RLVR. The majority are based on software engineering tasks from real-world repositories, while the rest are terminal-based agentic tasks sourced from open-source and internal datasets. In total, we used over 70k verifiable tasks across ~5k unique repositories. We deduplicate our environments against the repository sources from SWE-Bench and SWE-Bench-Pro to avoid source leakage during evaluation.

We used 64K and 128K context lengths for the first and second stages of SFT, respectively. This "long-to-longer" cascade approach enables bipartite training on valuable shorter data, establishing a robust performance baseline, followed by targeted long-context training only on high-quality verified samples.

---

## 📊 ZUSAMMENFASSUNG DER THEMEN DES TAGES

| Thema | Relevanz |
|-------|----------|
| **Anthropic Fable 5 Safety-Kontroverse** | Größte Story des Tages – massive Community-Reaktion, Anthropic rudert zurück |
| **Google DiffusionGemma** | Neues Open-Source-Paradigma: Diffusion für Text – Consumer-Hardware-tauglich |
| **Claude Fable 5 Release** | Neues Frontier-Modell, Benchmarks durchbrochen, aber stark gefiltert |
| **OpenAI IPO-Verschiebung + Modell 5.6** | Strategische Neuausrichtung + neues Modell in Pipeline |
| **Cohere North Mini Code** | Open-Source Coding-Agent-Modell, schlägt größere Modelle |

---

*Quellen: Wired, The Decoder, HuggingFace Blog, Simon Willison's Blog, Hacker News (Algolia API), The Information, Anthropic Research Blog*
*Recherche-Zeitraum: 10.–11. Juni 2026 | Erstellt: 11. Juni 2026*