🌁#92: Fight for Developers and the Year of Orchestration

Quite balanced report “Future of AI Research” with such contributors as Stuart Russell, Open Etzioni, Peter Norvig, Yoshua Bengio etc (pdf)

When AI met venture capital by Azeem Azhar

Is OpenAI's new story-generating model good at writing? And some AGI talk by Max Read

MCP is pretty cool, it actually changes the dev experience in a great way (your guide to MCP is here).

CoreWeave Strikes Gold with OpenAI

CoreWeave just landed a massive deal to supply AI infrastructure for OpenAI, securing up to $11.9 billion in contract value. The cherry on top? OpenAI is also taking a $350 million equity stake in CoreWeave. With Microsoft, Oracle, and now CoreWeave in its compute arsenal, OpenAI is stacking up allies for the AI arms race.

Cerebras Goes Big on AI Inference

Cerebras Systems is scaling up – fast. With six new AI inference datacenters across North America and Europe, the company is positioning itself as the largest domestic high-speed inference cloud. These facilities will crank out 40 million Llama 70B tokens per second, giving the industry a serious speed boost. With OpenAI’s o3 and DeepSeek R1 models demanding faster responses, Cerebras is betting big on real-time AI dominance. They also:

Little Gemma 3 that you can run on a single GPU or TPU is better than big Gemini – Researchers from Google DeepMind introduced Gemma 3, a lightweight, state-of-the-art open AI model optimized for single-GPU/TPU execution. It supports 140 languages, a 128K-token context window, and advanced text-visual reasoning. ShieldGemma 2, a 4B image safety checker, enhances AI safety →read more on their blog

Command A from Cohere unveiled a highly efficient generative AI model optimized for enterprise use. It matches or outperforms GPT-4o and DeepSeek-V3 in business, STEM, and coding tasks while requiring only two GPUs instead of up to 32. Command A processes 156 tokens/sec—1.75x faster than GPT-4o—and supports 256K context length →read more on their blog

OLMo 2 and Building Effective Teams for Training Language Models – Researchers from AI2 released open-source LLMs with 7B and 13B parameters, trained on 4T and 5T tokens, respectively. OLMo 2 Instruct outperforms Llama 3.1 8B Instruct and Qwen 2.5 Instruct. The team used Tulu 3’s post-training recipe, including RLVR, boosting scores by 4+ points. They emphasize FLOP efficiency, prioritization in training, and RL finetuning stability, pushing open-source AI capabilities further →read more on Interconnects

Baidu Unveils ERNIE 4.5 and Reasoning Model ERNIE X1, Makes ERNIE Bot Free Ahead of Schedule. ERNIE 4.5 is a multimodal foundation model, and ERNIE X1, a deep-thinking reasoning model. ERNIE 4.5 outperforms GPT-4.5 at 1% of its cost, with input/output priced at RMB 0.004/0.016 per 1,000 tokens. ERNIE X1, excelling in reasoning and tool use, costs RMB 0.002/0.008 per 1,000 tokens. Both models are free for individual users, with enterprise access via Baidu AI Cloud’s Qianfan platform →try on their website

Open-sourced MM-Eureka – researchers from Shanghai AI Laboratory developed a multimodal reasoning model applying large-scale rule-based RL to image-text tasks. It achieves stable accuracy gains, response length growth, and emergent reflection behaviors (visual "aha moments"). The model, trained on 54K samples, outperforms methods using 1M+ data, showing superior data efficiency →read the paper

Sesame AI Labs open-sources its conversational speech model CSM 1B. The model generates HQ speech from text using an Llama-based backbone and an audio decoder producing Mimi audio codes. It supports context-based speech generation and can process speaker turns. It lacks predefined voices and fine-tuning for specific ones. Ethical restrictions prohibit impersonation and misinformation. The model has 1.55B parameters, supports English, and is open-source for research and educational purposes →check on HF

Charting and Navigating Hugging Face’s Model Atlas by School of Computer Science and Engineering The Hebrew University of Jerusalem, Israel

Click to read the paper

🌟 Transformers without Normalization
Replace normalization layers with a lightweight transformation, improving training and inference speed while maintaining model accuracy.

SEAP: Training-free Sparse Expert Activation Pruning
Reduce computational costs by selectively activating relevant model parameters, improving efficiency while maintaining performance.

DistillM-2: A Contrastive Approach Boosts the Distillation of LLMs
Optimize knowledge distillation through contrastive loss functions, improving LLM preference alignment and decoding efficiency.

OmniMamba: Efficient and Unified Multimodal Understanding
Improve multimodal model efficiency by leveraging state-space models, reducing memory costs while maintaining high performance.

🌟 Communication-Efficient Language Model Training Scales Reliably
Reduce communication overhead in distributed training, making large-scale LLM training more efficient.

🌟 Monitoring Reasoning Models for Misbehavior
Detect hidden misalignment in LLMs by auditing their reasoning steps, highlighting challenges in ensuring transparency.

🌟 LMM-R1: Empowering 3B LMMs with Strong Reasoning
Improve multimodal reasoning through a two-stage reinforcement learning framework, enhancing both text and vision-based tasks.

Plan-and-Act: Improving Planning of Agents
Separate planning and execution in LLM agents, improving long-horizon task performance with structured synthetic data.

Gtr: Guided Thought Reinforcement Prevents Thought Collapse
Strengthen RL-based vision-language models by preventing loss of reasoning diversity, improving structured problem-solving.

Implicit Reasoning in Transformers is Reasoning through Shortcuts
Reveal that transformers often rely on statistical shortcuts rather than genuine multi-step reasoning.

Unified Reward Model for Multimodal Understanding
Introduce a reward model that evaluates both image and video tasks, improving preference alignment in multimodal models.

VisualPRM: An Effective Process Reward Model
Enhance reasoning in multimodal models using a reward-based approach, improving structured task completion.

Taking Notes Brings Focus? Multi-Turn Multimodal Dialogue
Improve multi-turn dialogue tracking by incorporating visual memory and stepwise reasoning modules.

SegAgent: Exploring Pixel Understanding
Develop an LLM-based segmentation model that imitates human annotators for pixel-level understanding.

MM-Eureka: Exploring Visual Aha Moment
Train multimodal models to develop "aha moments" in reasoning, improving math and vision-based tasks.

World Modeling Makes a Better Planner
Improve embodied AI planning by jointly optimizing state prediction and action selection.

MRT: Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
Reduce LLM response length without sacrificing accuracy, improving test-time efficiency.

FedRand: Enhancing Privacy in Federated Learning
Improve federated learning security by selectively updating model parameters, reducing data leakage risks.

Exploiting Instruction-Following Retrievers for Malicious Information Retrieval
Analyze how LLM retrievers can be manipulated to fulfill harmful queries, raising safety concerns.

Exploring the Vulnerabilities of Federated Learning
Examine gradient inversion attacks in federated learning and propose defenses to mitigate security risks.

Group-robust Machine Unlearning
Enhance fairness in machine unlearning by minimizing mutual information between model features and sensitive attributes.

Search-R1: Training LLMs to Reason with Search
Train LLMs to autonomously query search engines, improving retrieval-augmented reasoning.

New Trends for Modern Machine Translation
Redefine machine translation by treating it as a reasoning task, improving contextual and discourse-level accuracy.

Gemini Embedding: Generalizable Embeddings
Enhance multilingual text embedding models, improving performance across retrieval, clustering, and classification tasks.

Block Diffusion: Interpolating Between Autoregressive and Diffusion Models
Merge autoregressive and diffusion approaches, reducing generation steps while maintaining sample quality.

Sana-Sprint: One-Step Diffusion for Text-to-Image
Accelerate text-to-image diffusion models, reducing inference time while maintaining image fidelity.

CoRe2: Collect, Reflect and Refine for Text-to-Image
Improve text-to-image model efficiency through multi-stage inference, reducing computational costs.

Auditing Language Models for Hidden Objectives
Investigate how LLMs can develop covert misaligned goals, emphasizing the need for better alignment audits.

Can Large Reasoning Models do Analogical Reasoning?
Assess LLMs' ability to reason under perceptual uncertainty, highlighting weaknesses in analogical reasoning.

API Agents vs. GUI Agents: Divergence and Convergence
Compare API-based and GUI-based AI agents, discussing their strengths and future convergence.