Nov 10th ~ Nov 17th
#134 Latest AI Research Explained Simply

🗞️ Industry News in 1 Line

  1. ♥ 3.5k Liquid AI has released d1.2 and d1.2-ultra, omnimodal decision models that process text, images, and audio across five new API tasks, including value extraction, text span detection, bounding-box object localization, numerical range estimation, and multi-option selection. The d1.2-ultra model leads on quality while d1.2 provides a lower-cost option, with both offering calibrated probability scores and input-only pricing ($0.02 and $0.042 per million, respectively). You can try them today on the Liquid API via console.liquid.ai, with OpenRouter support and an open-source release planned for the future.

  2. ♥ 15k Alibaba has released Qwen4-Max, a MoE model with native vision and a 1M-token context window designed for tool use and long-horizon workflows across domains such as 3D modeling, manufacturing, finance, and scientific discovery. In internal evaluations, Qwen4-Max outperforms Qwen3.8-Max while using up to ~50% fewer tokens and reducing average per-task costs by over 60%. You can try it on Qwen Studio, QwenWork, Qoder, or the QwenCloud API, while open weights for smaller Qwen4 models are slated for future release.

  3. ♥ 10k Google has introduced Gemini Decision 1, a new decision model in the Gemini API with a 256K context window, 1K questions, and native support for image, audio, video, and PDF inputs. It is priced at $0.10 per 1 million input tokens with no additional costs. You can try it for free via Google AI Studio and the Gemini API.

  4. ♥ 5k MiniMax has released MiniMax H3-Image, an open-weight image generation model capable of producing 4K (16-megapixel) images in 10 seconds. The model supports prompt-based editing and up to 16 reference images, with hosted endpoints offering 50% off 16MP+ generations during launch week. You can try it on Hugging Face or via the MiniMax API, MiniMax Code, and Hailuo AI.

In partnership with

See the GTM Stack Behind Pebl’s Handoff

In global employment, the sale is only the start. Customers stay or leave based on what happens after signature.

At The GTM Garage, Pebl’s Chief Business Officer, Dutta Satadip, walks through the tools behind prospecting, pipeline, and handoff to onboarding, and how Pebl uses AI across the customer lifecycle.

15 min conversation. 15 min open Q&A. For Seed to Series A teams.

An Energy-Based Perspective on Machine Learning

Dai and Du [Georgia Institute of Technology, Harvard University]

Machine Learning

We divide ML into separate fields like supervised prediction, generative modeling, representation learning, and decision making, each relying on its own specialized rules. Because these silos speak different mathematical languages, engineers struggle to compare techniques or combine learned skills for new tasks without training fresh systems from scratch.

Logical composition of energy functions.

A tutorial from Georgia Tech offers a unifying fix by viewing these diverse methods through energy-based models, or EBMs. An EBM works like a scoring landscape that assigns a single energy number to any configuration of variables. Low energy marks plausible data, compatible features, or desirable outcomes, while higher energy signals unlikely configurations.

Instead of treating EBMs as one niche family, the authors show that nearly every major technique, including diffusion models, text representations, reinforcement learning, and solution verifiers, operates as an EBM. Each method simply makes different choices across three ingredients: defining the energy, searching or sampling for low-energy answers, and fitting that landscape to data.

Illustration of Negative Samples.

Organizing machine learning around these three choices clarifies why models trade off simple representations against costly sampling, while showing how researchers can invent algorithms by swapping one piece for another.

Since compatible energy functions measure suitability in the same way, engineers can add them together when running a model to satisfy multiple requirements simultaneously. For instance, a planner can combine an energy model of plausible robot trajectories with new terms for reaching a destination and avoiding obstacles, solving fresh tasks without retraining a dedicated predictor.

Unbundling Multimodal Vision Encoding

Brown et al. [Meta FAIR, New York University]

Multimodal models bycloud’s pick

Most multimodal AI models attach a separately trained vision encoder onto a language backbone. Since existing models vary in their data, architectures, and computing budgets, no one could isolate whether AI truly needs pretrained visual knowledge, dedicated image parameters, or a separate processing stage before text and images mix.

To solve this, the team built a controlled testbed where they trained every component from scratch on the same shared text and image data, by comparing models from 0.6 billion to 6.4 billion parameters at equal total compute.

A controlled study of multimodal vision encoding

These experiments proved that separate encoder pretraining is optional. An encoder trained jointly alongside the backbone matches separately pretrained encoders on visual question answering, even when they treat pretraining compute as free.

However, dedicated vision capacity is essential. Removing the encoder altogether leaves models trailing on visual understanding, particularly when reading text and charts. Simply adding the encoder's parameter count into the shared backbone fails to close this gap, because visual data needs its own dedicated parameters rather than competing with text.

Encoder-free trails the encoder on Avg. VQA, while Avg. Text stays comparable.

When the team placed small, vision-only modules directly inside the backbone's earliest layers, this integrated design matched a separate encoder at every scale and held up when starting from a pretrained language model.

Latent Feedback Is a Compute and Data Amplifier

Huang et al. [Institute of Foundation Models, USC, CMU]

LLM Feedback

Getting high-quality training text has become the main bottleneck for improving large language models. When fresh data runs out, developers usually train models on the same text repeatedly, even though each repeat delivers diminishing returns.

Researchers have found a clever way around this limit using latent feedback, a technique that makes every calculation and token count for more. Instead of spending extra compute on endless data repeats, a latent feedback Transformer takes each token's final state from the top layer of the network and feeds it directly into the input for the next token.

Latent feedback saves compute and unique data.

Because this step uses about 1.2 times more computing power per token, the model sees fewer data repetitions under the same compute budget. To test this tradeoff, the team compared latent feedback against standard Transformers across models from 60 million to 778 million parameters, tuning both systems across various data sizes and repeat counts.

The results show that once each parameter sees a modest baseline of unique text, a threshold modern models easily clear, latent feedback acts as a compute and data amplifier. With unique data fixed, it matches a standard Transformer's performance while saving up to 48 percent of the compute budget.

The data multiplier μ across unique data and model size.

With training passes fixed, it reaches the same performance using up to 47 percent less unique data, and downstream tasks confirm these gains. Best of all, it serves users with the same parameter count and memory cache as the standard model. By fitting a single scaling law to models up to 271 million parameters, the researchers predicted the 778 million parameter models' loss within 0.3 percent.

Interpretable weight-space programming builds transparent transformers and brain encoding models

Mehta et al. [Microsoft Research, Johns Hopkins University, Columbia University, New York University]

Weight Space Programming

Highly accurate neural networks often act like opaque black boxes whose inner workings cannot be deciphered. Meanwhile, models built from clear, human-understandable rules usually perform worse, forcing researchers to choose between accuracy and insight.

To bridge this divide, researchers tested a promising method called interpretable weight-space programming, where an AI coding agent builds a neural network's weights directly from scratch. Instead of using standard training or optimization loops, the agent writes explicit, readable code, checks its score on sample problems, and refines its logic over repeated iterations while a separate auditing agent blocks any hidden shortcuts.

On thirteen algorithmic puzzles like sorting and finding greatest common divisors, the strongest agent reached perfect accuracy on unseen tests by translating known algorithms straight into network weights, then trimmed those models down to a median of just 166 parameters. On trickier language tasks, the agent built simple, readable rules that worked well on three out of four datasets, and fresh AI sessions rebuilt nearly identical models using only descriptions of 200 words or fewer.

When researchers applied this approach to neuroscience, the agents built strong brain-prediction models and even caught hidden flaws in scientific data. For visual brain scans, an agent combined fourteen explicit image computations, such as local shading and bending contours, to beat traditional interpretable models and generalize to completely new people and datasets.

On electrical brain recordings of natural audio, its custom sound features outperformed textbook baselines, nearly matched the black-box speech model Whisper, and explained brain activity that Whisper missed.

Reply

Avatar

or to participate