MarkTechPost
·
1 week ago
● 4 sources
Thinking Machines Lab released Inkling, a 975-billion-parameter open-weights multimodal mixture-of-experts model with 41 billion active parameters and a 1-million-token context window. The model was trained on 45 trillion tokens of text, images, audio, and video, and features a controllable thinking-effort mechanism that allows users to trade inference cost for performance, achieving Terminal Bench 2.1 parity with Nemotron 3 Ultra using one-third the tokens. Inkling is available on Hugging Face and hosted platforms, enabling deployment of cost-tuned agentic systems and multimodal applications across voice, vision, and text inputs.
MarkTechPost
·
1 week ago
A German research consortium released Soofi S 30B-A3B, an open-source foundation model combining Mamba and Transformer architectures optimized for German and English. The model contains 31.6 billion parameters but activates only 3.2 billion per token, trained on 26.68 trillion tokens with German representing 15.32% of high-quality training data. Soofi S achieved the highest scores among fully open base models tested, scoring 70.1% on English benchmarks and 79.1% on German benchmarks, making it suitable for German document processing, bilingual code assistance, and long-context serving tasks.
Wired AI
·
1 week ago
Researchers at Meta, Stanford, and other institutions created the EgoBabyVLM Challenge, a test that measures how well vision language models can learn from approximately one thousand hours of egocentric video recorded from cameras worn by infants. Current cutting-edge AI models fail substantially on this benchmark, struggling to extract meaning from the messy, realistic footage that babies process efficiently. The findings suggest that designing AI systems with learning mechanisms inspired by infant brains—such as better attention mechanisms and social cue interpretation—could create more efficient models that learn from less data and require less energy.
Google Research
·
1 week ago
Researchers studied how diffusion models generate novel images and data beyond their training examples, finding that the models' creativity stems from neural network regularization effects that blur the learned score function. In a one-dimensional example with two training points at +1 and -1, weight decay causes the neural network to learn a smoother approximation of the score function, allowing generated particles to settle between the training points rather than copying them exactly. This score smoothing mechanism enables diffusion models to balance realism with novelty by interpolating between training data points while maintaining the quality of high-dimensional data manifolds.
Wired AI
·
1 week ago
● 3 sources
Thinking Machines Lab, founded by OpenAI exiles including former CTO Mira Murati, released Inkling, an open-weight AI model trained on audio, video, and text inputs. Inkling contains 975 billion parameters and performs comparably to leading Chinese open-weight models while being cheaper to run than closed alternatives. The release positions Thinking Machines as a competitor in the AI market and supports its vision for decentralized AI development outside the control of a few dominant companies.
TechCrunch AI
·
1 week ago
● 3 sources
Thinking Machines Lab released Inkling, an open-weight AI model with 975 billion total parameters that developers can download and modify, departing from the closed models sold by OpenAI, Anthropic, and Google. The model uses about 41 billion parameters per task and requires a third as many tokens as Nvidia's Nemotron 3 Ultra to achieve equivalent coding performance, according to the company's benchmarks. The startup is positioning Inkling as a customizable foundation for enterprises to fine-tune through its Tinker platform rather than a finished product, betting that organization-specific AI will outperform general-purpose models.
IEEE Spectrum AI
·
1 week ago
● 2 sources
Researchers at MIT unearthed ELIZA's original source code from archives, revealing that the 1960s chatbot was not a simple pattern-matching program but a sophisticated platform capable of adopting multiple personas beyond its famous Doctor therapist script. The source code, published in January 1966 in Communications of the ACM, showed ELIZA could assume roles discussing math, poetry, color, and other topics through different scripts like Neweng for weather small talk. The discovery demonstrates that Weizenbaum's technical innovations were more advanced than previously documented and that design decisions about language and interaction continue to influence modern AI development.
IEEE Spectrum AI
·
1 week ago
South Korean researchers at Seoul National University and Hanyang University developed Generative SNUPI, a generative AI model that designs DNA origami structures by automatically generating DNA sequences that fold into user-requested shapes. The model uses diffusion techniques to determine how DNA strands should be sequenced so molecular forces cause them to self-assemble into forms like dogs, stars, and the Mona Lisa, eliminating the tedious manual design work traditionally required. This capability could accelerate DNA origami applications in drug delivery and immunotherapy by removing a major bottleneck in the design process.
TLDR Dev
·
1 week ago
A researcher proposes 'Guardian Angels'—personalized LLMs that emulate individual users' values and preferences to amplify productivity and provide security against AI-powered attacks, arguing current chatbots are fundamentally misaligned with users and designed for replacement rather than augmentation. The approach combines dynamic evaluation, active learning, and continuous user feedback to create AI agents that remain under human control and learn user-specific patterns, addressing the principal-agent problem by unifying principal and agent goals. This shifts work from 'what and how to do things' to 'what is worth doing,' enabling users to deploy multiple specialized agents for productivity and security while maintaining strategic oversight.
TheSequence
·
1 week ago
OpenAI audited SWE-Bench Pro, a coding evaluation benchmark, and found that approximately 30 percent of its 731 public tasks contain defects such as rejecting correct solutions or accepting incomplete ones. OpenAI's agent-assisted review labeled 27.4 percent of tasks as defective while independent software engineers identified 34.1 percent as problematic. OpenAI withdrew its earlier recommendation that the field adopt SWE-Bench Pro as a standard evaluation tool due to these validity issues.
Together AI
·
1 week ago
● 2 sources
Thinking Machines Lab released Inkling, a 975-billion-parameter mixture-of-experts model with 40B active parameters that accepts text, image, and audio inputs for multimodal reasoning tasks. Inkling is available on Together AI's inference platform starting today with a 1M token context window and adjustable inference effort settings. Developers can now access a unified multimodal model through a single API endpoint that supports reasoning, coding, forecasting, and agentic workflows without managing their own infrastructure.
Apple ML Research
·
1 week ago
CLaRa is a framework that improves retrieval-augmented generation by compressing documents and jointly optimizing retrieval and generation in a shared continuous space. The method achieves state-of-the-art performance on QA benchmarks at a compression rate of 16, meaning documents are reduced to 1/16th their original length while maintaining answer quality. This unified approach allows gradients to flow through both retrieval and generation modules, aligning document relevance with answer quality during training.
Apple ML Research
·
1 week ago
Researchers propose FAE, a framework that adapts pre-trained visual encoders for image generation by using a single attention layer to convert high-dimensional features into low-dimensional latents suitable for generative models. On ImageNet 256×256, FAE achieves an FID score of 1.48 without classifier-free guidance after 800 epochs and 2.08 after 80 epochs. The approach enables simpler adaptation of pre-trained representations across different generative model families including diffusion models and normalizing flows.
Hugging Face Blog
·
1 week ago
Hume AI released Real World VoiceEQ, a benchmark that evaluates voice AI systems on human-quality interactions beyond traditional metrics like word error rates and latency. The benchmark drew from over 1 million human ratings across 40+ voice models and 60+ metrics spanning speech recognition, synthesis, and understanding tasks. The findings show no single voice model excels across all capabilities, and current systems struggle to interpret emotional cues and acoustic information that humans naturally recognize in conversation.
Hugging Face Blog
·
1 week ago
● 2 sources
Thinking Machines released Inkling, an open-source multimodal language model with 1 trillion parameters that accepts images, text, and audio inputs natively. The model has 975 billion total parameters with 41 billion active at any time, supports a 1 million token context window, and was trained on 45 trillion tokens across multiple modalities. Inkling is available on Hugging Face with immediate support in transformers, SGLang, vLLM, and llama.cpp, enabling developers to build multimodal reasoning applications through fine-tuning or remote inference.
