Swiss-AI Releases Apertus v1.5, a Fully Open Multimodal Model
Swiss-AI, an AI research institute based in Switzerland, has released **v1.5** (70B/8B), the latest version of its Apertus language model family, on Hugging Face. Its biggest distinguishing feature is that it releases **not only the model weights but also the training data and training recipe** in full.
Multimodal input**: Supports image and audio input in addition to text. Output is text.
Long context**: Handles up to 262,144 tokens (roughly 260K tokens) by default — 4x the previous limit.
Thinking Mode**: Includes a mode that reasons before answering on reasoning tasks.
Multilingual support**: Covers a wide range of languages.
Fully open**: Model weights, training data, and training methodology are all released. Designed to respect data owners' opt-outs.
Apertus 1.5 was built by continuing pretraining from its predecessor 1.0 with an additional 4T tokens of multimodal mixture data for the 8B model and 2T tokens for the 70B model. It is a decoder-only Transformer using the xIELU activation function and the AdEMAMix optimizer, and reportedly performs on par with other models of similar size. The technical report is scheduled for release within the next few weeks.
LLMs Can't Keep Up with "Evolving User Intent"
"LLMs Get Lost in Evolving User Intent," a study that drew attention on Hugging Face Daily Papers, highlighted the phenomenon that strong performance in static single-turn settings **does not transfer to dynamic multi-turn conversations**.
In real conversations, users rarely state their full intent upfront; they gradually reveal, revise, or pivot their intent as the conversation unfolds. The research team introduced a framework that converts existing static tasks into dynamic multi-turn conversations in which "user intent evolves across turns," and validated it on multiple benchmarks.
The results consistently showed that, regardless of model family, **high scores in static settings do not transfer to dynamic settings**. According to the team, this is an "invisible gap" for future collaborative AI agents — a serious shortcoming that static evaluations cannot detect.
A Learning Method That "Distills" Agent Experience
Another notable paper, "Sample-Efficient Learning from Agent Experience," proposes a method named **Experience Distillation**. It internalizes, into the model weights, the experience an agent gains through environment interactions — without any further interaction with the environment.
In experiments, the method was validated on 749 software engineering tasks and 6 text adventure games. It **retained at least 64.8% of the in-context learning gains**, whereas conventional supervised fine-tuning recovered only 3.8%. It also matched the performance of reinforcement learning baselines with **9.6x fewer environment samples**.
This looks like a promising approach for real-world agent learning, where costly environment interactions are the bottleneck.
MIT Media Lab: AI Video Undermines "Trust in Real Footage"
"Seeing Is Not Believing," a study from the MIT Media Lab, argues that realistic AI-generated video is **shaking people's confidence in real footage itself**.
As AI-generated video quality improves, even genuine footage now invites the doubt "this might be AI-generated." The study analyzes the structural impact this phenomenon has on societal trust in visual evidence.
Going beyond the "deepfake problem" framing, it carries implications for law and journalism, in that **the evidentiary value of authentic footage itself declines**.
"License Laundering" in AI Data Supply Chains
"Don't Trust the Label: License Laundering in AI Supply Chains," published on arXiv, is a paper that flags the problem of **license misrepresentation (License Laundering)** in AI training data supply chains.
The authors dub "license laundering" the phenomenon whereby, as data passes through multiple intermediaries, the original license terms become unclear or the data gets relabeled under a more permissive license. This is a structural challenge that undermines transparency into the provenance and usage terms of AI models' training data, and it deserves attention from legal and compliance perspectives as well.
Adecco Report: AI Won't Collapse Employment
Adecco Group, one of the world's largest staffing firms, has published its view that **AI will not cause a collapse of employment** (reported by Reuters). The company tracks labor market trends on a global scale, and its analysis is that while AI replaces some tasks, it also creates new job categories and demand.
The topic has also drawn attention on X (formerly Twitter) as a debate over "why AI isn't driving up unemployment," with the prevailing view that AI's actual impact on employment will be limited in the short term.
Palo Alto Networks: Generative AI Makes Cyberattacks "Blazing Fast"
Seiryo Someya, Chief Cybersecurity Strategist at Palo Alto Networks, warned in a talk at SoftBank World 2026 that the spread of generative AI is **shortening attacks "from 2 days to 25 minutes"** (reported by ITmedia).
Attack code generation and phishing lure writing powered by generative AI are becoming more sophisticated and faster, leaving companies increasingly unable to cope with traditional defenses alone. He stressed that defenders' response capabilities, also leveraging AI, must be strengthened in parallel.
---