AI & Tech

What happened

Google DeepMind has announced a new offering called Gemini 3.5 Transcribe, a speech-to-text solution that the company says delivers more intelligent transcription.

The announcement frames the tool as an advancement in converting spoken words into written text, with a focus on improved intelligence in the transcription process.

Why it matters

The emphasis on 'intelligent' transcription suggests a shift beyond simple word-for-word conversion, potentially involving deeper understanding of context, nuance, and speaker intent. If realized, this could make automated transcription significantly more useful in complex real-world settings.

Key facts

Google DeepMind announced Gemini 3.5 Transcribe.

Gemini 3.5 Transcribe is described as providing more intelligent speech-to-text transcription.

What to watch next

More details about the features, availability, and underlying technology of Gemini 3.5 Transcribe are expected to follow from Google DeepMind.

Observers will be watching for how 'intelligent' transcription translates into practical performance and whether this sets a new benchmark for speech recognition tools.

Sources

• Google DeepMind

What happened

A new AI-driven tool called CrysVCD, developed at MIT, is designed to help create materials that perform well in practical applications.

The tool targets a major bottleneck in materials development: the lengthy and expensive process of eliminating chemically unstable designs.

Why it matters

If CrysVCD delivers on its promise, it could significantly lower the time and financial resources needed to bring new materials from concept to real-world use.

More efficient screening could open the door to faster innovation across fields that depend on advanced materials.

Key facts

CrysVCD is a tool developed at MIT.

The tool uses AI to help design new materials that work in the real world.

It could cut the huge amounts of time and money spent on screening out chemically unstable designs.

What to watch next

Look for further updates on how CrysVCD is applied in materials research and whether it gains wider adoption.

Sources

• MIT News

What happened

Apple Machine Learning presented Luce, a 3D representation designed for high-fidelity image-to-3D generation that captures both geometry and appearance.

The representation combines geometry with physically based rendering (PBR) materials—including albedo, metallic-roughness, and surface normals—to support relighting and integration into standard rendering pipelines.

Luce organizes geometry and PBR materials into a voxelized multimodal Gaussian cloud, where each modality gets its own dedicated Gaussian primitives, and a variational autoencoder compresses it into a unified material-aware latent space.

Why it matters

By incorporating PBR materials directly into the 3D representation, Luce aims to make generated assets compatible with standard rendering pipelines, which could make image-to-3D outputs more usable in real-world workflows that require relighting.

The unified material-aware latent space suggests a path toward generating geometry and appearance together rather than as separate, less coherent components.

The use of multimodal Gaussian primitives is a step toward representations that are both expressive enough for detailed assets and structured enough for efficient generation.

Key facts

Luce is a 3D representation that unifies geometry and PBR materials in a voxelized multimodal Gaussian cloud.

Dedicated Gaussian primitives are used for each modality, including albedo, metallic-roughness, and surface normals.

A variational autoencoder compresses the representation into a unified material-aware latent space.

What to watch next

The paper description is partial, so forthcoming details may clarify how Luce is used for actual image-to-3D generation and rendering.

Keep an eye on Apple's machine learning publications for experimental results and comparisons with other 3D generation methods.

Sources

• Apple Machine Learning

What happened

Distilling tool-calling abilities typically starts with supervised fine-tuning on teacher-generated trajectories. Production pipelines repeat this step on a daily or weekly basis, each time paying the cost of a frontier teacher model.

The standard mechanism is simple: keep the teacher’s successful trajectories and discard the rest. Because failures provide no signal, the same hard scenarios tend to reappear cycle after cycle.

On the τ2-bench evaluation, 57% of teacher trials fail, and two-thirds of those failures are near-misses where most tool calls were correct but the trajectory stalled unfinished.

Why it matters

Re-running the same generate-and-filter loop means teams repeatedly pay frontier-teacher fees without ever learning from what went wrong. The hard scenarios that block progress stay unsolved.

Near-misses contain valuable signal: they show a model was close to the right behavior. Ignoring them may be leaving cheap improvements on the table, which is exactly the gap PROOF-Gen appears to target by connecting optimized data to better distillation.

Key facts

Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities.

Post-training pipelines re-run distillation on a daily or weekly cadence, paying frontier-teacher cost each cycle.

The current mechanism is generate-and-filter: keep passing trajectories, discard failures.

On τ2-bench, 57% of teacher trials fail, two-thirds of them near-misses with most tool calls correct.

What to watch next

Whether PROOF-Gen’s use of optimized data—potentially including failed or near-miss trajectories—can break the generate-and-filter loop and improve distillation efficiency.

If successful, this approach could reduce the recurring cost of frontier-teacher calls and help models finally clear persistent hard scenarios.

Sources

• Apple Machine Learning

What happened

A new paper from Apple Machine Learning introduces IDEA Prune, a pipeline that integrates enlarged model pretraining with structured pruning during generative language model pretraining.

The authors argue that previous pruning work often ignores the role of enlarged pretraining, and they propose studying enlarge-and-prune as a single system.

The research addresses whether pretraining an enlarged model is worthwhile even when that model is never deployed, and how to optimize the integrated pipeline.

Why it matters

As large language models grow, staying within limited inference budgets becomes critical, and this work reframes pruning as part of pretraining rather than a post-hoc step.

If structured pruning pipelines can deliver token efficiency over training target-size models from scratch, they could reduce the cost of producing deployable models.

Key facts

The pipeline is called IDEA Prune and combines enlarge-and-prune with generative language model pretraining.

Structured pruning pipelines have shown promise in token efficiency compared to training target-size models from scratch.

The paper was published by Apple Machine Learning on August 26, 2026.

What to watch next

Future details on how the pipeline optimizes the trade-off between enlarged pretraining cost and downstream deployability.

Whether this integrated approach influences how other labs design pretraining and pruning strategies for language models.

Sources

• Apple Machine Learning

What happened

Google Research has unveiled AgentHands, a system designed to generate interactive hand gestures for spatially grounded agent conversations in extended reality (XR).

The work, categorized under Human-Computer Interaction and Visualization, aims to enhance the realism and engagement of virtual agents by enabling them to use natural hand movements during conversations.

Why it matters

In XR environments, realistic non-verbal communication is crucial for creating immersive and believable interactions. AgentHands addresses a gap by focusing on gestures that are spatially grounded, meaning they relate to the physical or virtual space around the agent.

This development could significantly improve user experience in applications like virtual assistants, training simulations, and collaborative workspaces, where natural communication is key.

Key facts

AgentHands generates interactive hand gestures for agent conversations in XR.

The research is from Google Research and falls under Human-Computer Interaction and Visualization.

The system emphasizes spatially grounded gestures, aligning with the XR context.

What to watch next

Future integration of AgentHands into commercial XR platforms or development tools.

Potential extensions to full-body gestures or emotional expressions to further enhance agent realism.

Sources

• Google Research

What happened

Apple Machine Learning researchers introduced STARFlow2, a framework that bridges language models and normalizing flows for unified multimodal generation.

The approach addresses structural fragmentation in existing multimodal models, which often sacrifice visual fidelity, impose asymmetry, or degrade pretrained understanding.

The key insight is that autoregressive normalizing flows share the same causal mask, KV-cache mechanism, and left-to-right structure as LLMs.

Why it matters

This work could lead to more coherent and efficient multimodal systems that handle text and images in a unified manner without compromising quality.

By aligning normalizing flows with LLM architecture, it may enable seamless integration of generation capabilities into existing language model frameworks.

The approach has potential to improve visual fidelity in generated images while maintaining strong text understanding.

Key facts

STARFlow2 is a unified multimodal model for understanding, reasoning over, and generating interleaved text-image sequences.

Existing approaches face issues like discrete tokenization reducing visual fidelity, structural asymmetry from combining causal text generation with diffusion denoising, and degradation of pretrained understanding in vision-language models.

Autoregressive normalizing flows are structurally identical to autoregressive Transformers, sharing causal mask, KV-cache, and left-to-right generation.

What to watch next

Future benchmarks will reveal how STARFlow2 compares to existing multimodal models in terms of generation quality and reasoning performance.

Watch for potential applications in tasks requiring seamless text-image interleaving, such as visual storytelling or multimodal dialogue.

The framework may inspire further research into unifying generation paradigms with LLM architectures.

Sources

• Apple Machine Learning

What happened

On 24 August 2026 Meta published MetaRoCE, an RDMA transport protocol purpose-built for AI workloads running over commodity Ethernet.

Released alongside it: the MetaRoCE specification, a reference software implementation, and a compliance test.

Why it matters

Training and serving frontier models depend on moving data between GPUs quickly and reliably; time spent on that is compute sitting idle. Transport-level design decides how much of it is wasted.

Shipping a specification, a reference implementation and a compliance test together — rather than a paper alone — reads as an invitation for others to implement against it.

Key facts

Name: MetaRoCE, an RDMA transport for AI workloads.

Environment: commodity Ethernet.

Approach: clean-sheet rather than an extension of an existing protocol.

Published: specification, reference software implementation, compliance test.

Source: Meta Engineering blog, 24 August 2026.

What to watch next

Whether NIC and switch vendors implement it — the usual gate for a transport protocol.

Migration paths from existing RoCE deployments.

Sources

• Meta Engineering

What happened

Researchers have developed a novel algorithm that can generate scenarios for extreme events, even when historical data on such events is scarce or nonexistent.

The algorithm is designed to anticipate the unprecedented scenarios that critical infrastructure and global supply chains are least prepared for, according to the source.

Why it matters

Traditional risk assessment often relies on past data, which may not capture the full range of possible extreme events. This new approach could help organizations better prepare for rare but impactful disruptions.

By learning to anticipate unprecedented scenarios, the algorithm could enhance resilience in sectors that are vital to modern society, such as energy, transportation, and logistics.

Key facts

The algorithm learns to anticipate unprecedented scenarios.

It targets critical infrastructure and global supply chains.

It does not require extreme data to generate extreme event scenarios.

What to watch next

Watch for further details on how the algorithm is trained and validated, as well as potential pilot applications in real-world infrastructure systems.

Also watch for discussions on the algorithm's limitations and how it might be integrated into existing risk management frameworks.

Sources

• MIT News

What happened

Multimodal large language models often rely on visual chain-of-thought (Visual CoT), which generates intermediate reasoning images to handle spatial, temporal, and embodied tasks.

While Visual CoT aids visual foresight, it adds significant inference overhead, especially for proactive video reasoning.

Apple's new framework, Internalized Visual Thinking (IVT), shifts visual reasoning to training time, allowing direct reasoning at inference without intermediate image generation.

Why it matters

This approach could make video reasoning more efficient and practical for real-time applications, reducing computational burden.

By internalizing visual thinking, models may achieve faster inference without sacrificing reasoning quality, potentially enabling broader deployment.

Key facts

Visual CoT generates intermediate reasoning images for spatial, temporal, and embodied reasoning.

Visual CoT introduces substantial inference overhead.

IVT is a post-training framework that jointly optimizes textual prediction and visual thinking.

IVT aims to learn visual thinking during training and reason directly at inference.

What to watch next

Whether IVT can maintain or improve reasoning accuracy compared to Visual CoT in real-world video tasks.

How IVT's training efficiency and scalability compare to existing methods in large-scale deployments.

Sources

• Apple Machine Learning

What happened

Google DeepMind has announced a partnership with game studios to prototype what it describes as breakthrough AI-driven gameplay.

The effort builds on 15 years of AI research in games, drawing on a lineage that spans from early Atari experiments to the multiplayer universe of EVE Online.

Why it matters

This move signals a shift from using games purely as research benchmarks toward embedding AI directly into player-facing experiences.

By collaborating with studios early in the prototyping stage, DeepMind may be aiming to shape how AI is woven into game design rather than bolted on afterward.

Key facts

Google DeepMind is partnering with game studios.

The partnership aims to prototype breakthrough AI gameplay.

The project draws on 15 years of AI research in games, from Atari to EVE Online.

What to watch next

Which game studios are involved and how their prototypes use AI in actual gameplay will be key signals.

It remains to be seen whether this partnership leads to commercial releases or stays in the research prototype phase.

Sources

• Google DeepMind

What happened

New research from MIT is exploring ways to improve the materials used in producing ammonia without relying on fossil fuels.

The work focuses on enabling a cleaner process for making this essential chemical, which is widely used in fertilizer and other products.

Why it matters

Ammonia is critical to global agriculture, but conventional production methods depend heavily on fossil fuels.

Developing better materials for a fossil-fuel-free process could significantly reduce the environmental footprint of producing this essential chemical.

Key facts

The research comes from MIT.

The goal is to find better materials for fossil-fuel-free ammonia production.

Ammonia is essential to fertilizer and other products.

What to watch next

Whether these material discoveries can be developed into practical, scalable technologies for industrial ammonia production.

Future research may identify specific materials that make the fossil-fuel-free process efficient enough for real-world use.

Sources

• MIT News

What happened

Microsoft Research has released Skala 1.1, an updated version of its deep-learning exchange-correlation functional designed to improve predictions in density functional theory (DFT).

The update delivers greater accuracy and expands accessibility across the computational chemistry ecosystem, with a living benchmark to track computational performance over time.

The release reflects an effort to create a faster path to predictive DFT by broadening who can use the tool.

Why it matters

DFT is widely used in chemistry and materials science, but accuracy and ease of use have historically been limits; a more accurate and accessible deep-learning functional could lower barriers for researchers.

The inclusion of a living benchmark suggests a commitment to ongoing performance tracking, which may help the community evaluate progress transparently.

Expanding access across the computational chemistry ecosystem could speed adoption and push predictive modeling forward in practical settings.

Key facts

Skala 1.1 is an updated deep-learning exchange-correlation functional from Microsoft Research.

The update provides greater accuracy and expanded accessibility across the computational chemistry ecosystem.

It includes a living benchmark to track computational performance.

What to watch next

Whether the broader availability of Skala 1.1 leads to wider integration into existing computational chemistry workflows.

How the living benchmark evolves and what it reveals about the functional's real-world performance as more researchers use it.

Sources

• Microsoft Research

What happened

The paper introduces an iterative pseudo-labeling training approach for Mandarin-English code-switching automatic speech recognition, marking the first application of this method to CS-ASR.

The approach uses a large unlabeled corpus to generate pseudo-labels, creating a semi-supervised dataset for training.

Training proceeds in three phases: pseudo-label generation, two-stage bilingual model training, and iterative improvements.

Why it matters

Code-switching, where speakers alternate languages within a single utterance, is notoriously difficult for ASR systems because dedicated training data is scarce.

By leveraging unlabeled data through iterative pseudo-labeling, this approach offers a potential path to improving CS-ASR without relying on costly manually transcribed code-switched speech.

The iterative refinement process could help models progressively improve their own predictions, making better use of available bilingual resources.

Key facts

Code-switching involves alternating languages within the same utterance and poses significant challenges for ASR.

This paper applies iterative pseudo-labeling to CS-ASR for the first time.

The approach has three phases: pseudo-label generation, two-stage bilingual model training, and iterative improvements.

What to watch next

Whether the iterative pseudo-labeling approach can be extended to other code-switching language pairs beyond Mandarin-English.

How the quality of generated pseudo-labels evolves across iterations and whether it leads to sustained performance gains.

Potential integration of this semi-supervised method into practical ASR systems deployed in multilingual or code-mixed contexts.

Sources

• Apple Machine Learning

What happened

Apple Machine Learning researchers are investigating how multilingual language models can acquire knowledge for low-resource target languages from high-resource languages.

The work centers on lexical interventions as a way to support tasks like scientific reasoning, commonsense inference, and world knowledge when target-language training data is limited.

Existing approaches to cross-lingual knowledge transfer typically require large amounts of parallel data, translation systems, auxiliary models, or additional training stages.

Why it matters

Many languages lack enough training data to build high-performing models on their own, so effective transfer from high-resource languages is essential for practical multilingual systems.

If lexical interventions can reduce the need for heavy parallel data and extra components, cross-lingual knowledge transfer could become more feasible for a wider range of languages.

Key facts

Cross-lingual knowledge transfer is critical for multilingual models serving languages with insufficient training data.

When target language data is scarce, knowledge comes primarily from the high-resource language.

Existing transfer improvement methods require large parallel data, translation systems, auxiliary models, or additional training stages.

What to watch next

Whether the lexical-intervention technique works across a broad set of downstream tasks and language pairs.

How the method compares with existing resource-heavy transfer approaches in real low-resource settings.

Sources

• Apple Machine Learning

What happened

As language models scale, they demand ever more training data, but some valuable sources — such as low-resource languages or specialized domains — exist only in limited amounts.

Researchers at Apple Machine Learning investigated a common workaround: blending that scarce, high-value target data with abundant generic data.

Using more than 2,000 language-model training runs, the study analyzed the balance between underexposing the model to the target domain and repeating the same target examples too often.

The findings show that too little target data weakens target-domain performance, while too much leads to diminishing returns and eventual overfitting.

Why it matters

This work highlights a central tension in pretraining: data scarcity cannot always be solved by adding more data, so mixture design becomes a critical control knob.

For builders of models aimed at specialized or underserved domains, understanding this trade-off is essential for choosing how aggressively to weight limited target data without sacrificing generalization.

Scaling laws that account for constrained data sources could lead to more principled decisions about dataset composition as model sizes grow.

Key facts

Language model scaling increases data requirements, yet some target data sources are inherently limited in size.

A common strategy is to mix scarce target data with abundant generic data.

The study examined this trade-off across more than 2,000 language-model training runs.

What to watch next

Further work may translate these findings into practical guidance on optimal mixture ratios for data-constrained domains.

Expect future scaling-law research to incorporate data repetition and overfitting effects explicitly when predicting model performance.

Sources

• Apple Machine Learning