Deepseek, GPT-Live-1 and more: Today's AI News
Deepseek saves memory for agents, OpenAI opens GPT-Live-1, DeepMind maps the genome, and Microsoft patches nearly 1000 vulnerabilities.
Inhaltsverzeichnis
Today brings several news items that show where AI is heading right now: away from pure demo glamour and toward production-ready infrastructure. Especially exciting are the advances in agents, language models, and scientific applications — exactly the areas where AI doesn’t just sound nice, but becomes measurably useful.
At the same time, we also see the downside of the boom: security vulnerabilities, copyright disputes, and the question of who actually pays for the music, data, and models. Welcome to an entirely normal AI Monday — just on a Friday.
🤖 Deepseek V4.1-Flash: Less memory, more agents
Deepseek is introducing V4.1-Flash, a new multimodal model that at first glance sounds like a typical “bigger, faster, better” release — but in practice is mainly more efficient. According to the report, the model reduces KV-cache memory to one quarter of the predecessor’s. That matters a lot for AI agents, because memory is often the hidden cost driver when models have to process longer contexts, tools, and multi-step tasks. Particularly interesting: although only 16 billion parameters are active per token, V4.1-Flash is said to land just ahead of Opus 5 and GPT-5.6 Sol on the coding benchmark DeepSWE. The model is released under the MIT license — a gift for anyone building open, more affordable agent stacks. Source: The Decoder
🎙️ GPT-Live-1: OpenAI brings full-duplex voice AI to the API
OpenAI is opening GPT-Live-1 to developers — a speech model that can listen and speak at the same time. That’s exactly the difference between “wait until you’re done” and “interrupts you like an overzealous call center agent.” In benchmarks, the model is 30 percentage points ahead of its predecessor, and the Speak learning app reports 80 percent fewer interruptions. It sounds modest, but for voice AI it’s a major step: more natural conversations, fewer pauses, better real-time interaction. The catch remains the price: $0.05 per minute is fine for prototypes, but for heavy users it quickly becomes a real line item. For products involving telephony, voice coaching, or support automation, this is still highly interesting. Source: The Decoder
🧬 AlphaGenome Atlas: DeepMind maps the genome at AI scale
With the AlphaGenome Atlas, Google DeepMind is showing how big AI in biology is thinking now: the dataset contains predictions for all roughly nine billion possible single-letter changes in the human genome. That’s not only academically impressive, but potentially medically relevant. DeepMind aims to estimate which DNA variants could cause disease — and in one epilepsy case, the atlas even helped identify a previously overlooked variant as the likely cause. The scale is enormous: one petabyte of data, more than 30 times the size of the AlphaFold database. For research and diagnostics, this signals that AI is no longer just folding proteins, but increasingly serving as a hypothesis engine for genomics. Source: The Decoder
🔐 Microsoft closes nearly 1000 security vulnerabilities
On Patch Tuesday, Microsoft closed a record number of vulnerabilities: nearly 1000 security holes, including two that are already being actively exploited. That’s not a pretty number, but unfortunately it’s a very real reminder that the AI revolution still rests on classic IT security. For companies, that means: patch, patch, patch — and not sometime later, but immediately for critical issues. Especially in AI environments with APIs, model serving, data pipelines, and cloud integrations, an unpatched vulnerability can quickly become an entry point. The AI hype changes nothing about the fact that poor update hygiene is still the most popular attack surface. Source: Heise
🎵 Universal Music launches licensed AI music platform with ElevenLabs
Universal Music Group is building a new AI music platform together with ElevenLabs, where users can access licensed material from the catalog and create remixes, mashups, and new versions from it. This is a notable move because it shows that major rights holders don’t just want to fight AI, but to monetize it in a controlled way. For the music industry, this could become a model for future licensing deals — less Wild West, more licensing and accounting logic. For creators, the question remains, of course: how fairly will rights, royalties, and opt-in/opt-out actually be handled? Still, this is an important signal that AI music is moving from the experimental phase into licensed product offerings. Source: The Verge
⚖️ Anthropic: $1.5 billion settlement — and now the payout chaos
Anthropic has accepted a $1.5 billion settlement in the copyright dispute, but according to the report there is now disagreement over how it should be distributed among authors and publishers. This is the largest copyright settlement in U.S. history so far — and another chapter in the question of who pays for training data and who ultimately gets compensated. This matters for the industry because such settlements can quickly become precedents: not only legally, but economically as well. If even the payout becomes difficult in the details, it shows one thing above all: the AI industry still lacks a clean, broadly accepted standard for compensation and rights clearance. In short: the technical race is fast, the legal department is running behind. Source: The Decoder
📐 GPT-6 Astra surprises in mathematics
OpenAI’s GPT-6 Astra takes first place in ErdosBench for open mathematics problems — even though mathematics was not explicitly prioritized, according to chief scientist Jakub Pachocki. That’s interesting because it fits a thesis we are seeing more and more: instead of becoming uniformly better at everything, frontier models are increasingly developing extreme strengths in specific domains. OpenAI is therefore investing more heavily in recursive self-improvement and alignment research, while certain capabilities more or less “come along for the ride.” For users, that means: even if a model isn’t specifically tuned for math, it can suddenly perform strongly there. For research, it’s a sign that capability development and research priorities are not always aligned — and that’s exactly what makes the current model generation so hard to predict. Source: The Decoder
🛠️ Tool tip of the day
If you want to experiment yourself with multimodal models, voice AI, or open agents, it’s worth taking a look at a flexible API and workflow environment that lets you quickly test and combine models. Especially with news as varied as Deepseek, GPT-Live-1, and the licensing topics around AI music, you can see it clearly: whoever can prototype quickly has the advantage. #
Don’t want to miss any news? Subscribe to the newsletter