This research investigates the transferability of supervised fine-tuning (SFT) lessons across alignment training, model organisms, and toy models. Through three experiments, the authors demonstrate that lessons in behavior generalization, capability preservation, and robustness can transfer across domains, suggesting researchers should borrow techniques from other areas.

Researchers propose modifications to encoder-decoder transformers for automated molecular structure elucidation from IR spectroscopy without requiring pre-determined chemical formulas. Using a Mixture-of-Experts (MoE) decoder with non-additive aggregation (linear-order statistics and Choquet integral) and an auxiliary contrastive alignment loss, they improve Top-K prediction accuracy by over 10 percentage points over baseline IR-only models. Analysis confirms IR spectra encode most relevant chemical information, broadening AI's utility in analytical chemistry.

A new framework called MeRLa meta-learns a task-aware shaping function to improve reward models in RLHF, leading to better alignment and performance. Experiments on LLaMA-3-8B show significant gains over existing methods like PPO and DPO, with reduced instability.

Sim2Win is a team-agnostic pre-match tactical recommendation framework that uses StatsBomb event data to build tactical profiles, clusters team behaviors into eight playstyles via K-Means, and trains thirteen classifiers to predict win/draw/loss probabilities. In LOCO evaluation, it achieves mean ROC-AUC of 0.704 and accuracy of 55.4% on unseen teams, outperforming ELO, Pi-Rating, and GAP baselines.

Researchers found that deep reinforcement learning agents with frozen random CNN feature extractors spontaneously develop sparse representations in the fully-connected layer, with only 1-3 out of 64 neurons active for deterministic Pong. The sparsity scales with task complexity, and removing active neurons crashes performance. The active set locks early in training, and the phenomenon reveals the effective rank of the underlying problem without explicit sparsity mechanisms.

Textideo's AI Facebook Ad Video Generator uses advanced AI models to help businesses and marketers quickly convert product images, descriptions, and ideas into professional video ads. No editing skills or expensive production teams required. It supports product promotions, brand stories, educational videos, and more.

Premation is an open-source AI alternative to After Effects, leveraging artificial intelligence for motion graphics and compositing.

The author argues against the AI-driven solutionism prevalent in Silicon Valley, using medicine as a case study to illustrate that intelligence is not the main bottleneck for real-world progress. Governance, regulation, and incentives matter more. The essay critiques the monoculture in San Francisco tech circles that dismisses these constraints.

A report reveals that Google AI Studio's delete function fails to actually remove user chat data from servers, as demonstrated in three videos, raising privacy concerns.

Kinney Drugs' AI assistant 'Burt' has led to prescription errors, delays, and privacy concerns, frustrating customers and overwhelming pharmacists. The rollout highlights gaps in data privacy regulations.

Databricks announces agentic code converter in Genie Code (Beta) that converts proprietary SQL dialects to open ANSI SQL using parallel subagents for iterative conversion, syntax and semantic validation. It simplifies data warehouse migration with project management, complexity scoring, lineage, and custom skills.

Databricks launches the agentic code converter (Beta) using Genie Code AI to automatically convert proprietary SQL dialects (T-SQL, Snowflake, Redshift, Oracle, BigQuery, Teradata) to open ANSI SQL, with features for migration projects, complexity scoring, lineage analysis, and custom skills, dramatically simplifying legacy data warehouse migration to the Lakehouse.

A video discusses whether the AI industry is facing a bubble burst, highlighting signs of overvaluation and speculative investment.

Claude Code Merge Queue is a local, zero-cost tool that serializes parallel agent landings to prevent push races and redundant builds. It integrates with Claude Code agents and provides a set of commands for managing lanes, builds, and deployments.

This article discusses how to package and sell AI skills as products, focusing on selling decision-making expertise rather than generic knowledge. It emphasizes identifying repetitive tasks, creating narrow specific skills, testing them, and pricing based on saved effort.

avatarin uses OpenAI's GPT-Realtime to give Yamada Denki shoppers 24/7 multilingual support. In two weeks, 30,000 people used the agent and 92% of survey responses were positive.

To operate effectively across contexts, robots must adapt their action execution. This paper proposes MoMo, a two-stage imitation learning framework with a spatiotemporal action tokenizer and a behavior-cloning transformer that takes a continuous motion-mode condition. In six real-robot tasks, varying this condition produces steady, dynamic, and intermediate behaviors distinguishable by human raters. MoMo transfers unseen motion modes to tasks demonstrated in only one mode, largely preserving task success, showing compositional generalization.

While UMAP is widely used for exploring high-dimensional data, typical workflows focus on its lower-dimensional embedding, largely overlooking the rich k-nearest-neighbor (kNN) graph that UMAP constructs internally. This graph encodes the data manifold in its original high-dimensional space, before the distortion that UMAP’s 2D projection introduces. We demonstrate the untapped potential of this internal representation, showing how standard graph algorithms applied to this graph enhance data sensemaking: (1) PageRank identifies representative data points, (2) k-core decomposition reveals dense core regions versus sparse periphery, and (3) clustering coefficient detects tight-knit neighborhoods with highly-similar data points. Through quantitative and qualitative evaluation on MNIST and Fa…

This article explores how AI can solve retrieval problems in creative workflows through semantic search and vector databases, rather than replacing creatives. The author argues that the main friction in creative work is organizing, finding, and reusing existing assets, and AI as an infrastructure layer can improve accessibility and reduce redundant effort.

The Generative AI Studio employs art and design methodologies—studio structure, critique, and charrette—to teach generative AI as a creative medium rather than a mere tool, fostering judgment and discernment in students.

Today's AI news focuses on AI's penetration into financial services, covering applications across subsectors. Also reports on Kimi K3 open model, OpenAI Codex security tools, and academic access. AIE NYC opens early bird tickets with finance theme.

OpenAI Group PBC today launched a program that will give 100,000 researchers free access to ChatGPT’s advanced features. Part of a $250 million initiative, the program initially opens to 10,000 users. Participants gain higher usage limits, larger context windows, and an expanded deep research feature. The program also includes Codex, ChatGPT Work, training, and integration support.

Microsoft reported $90B in Q4 FY2026 revenue (+18% YoY), beating estimates. Cloud revenue hit $59.3B (+27%), Azure up 43%, and M365 Copilot paid users grew 50% QoQ. Capex reached $41B (+70% YoY), but operating cash flow remained strong.

OpenLore is an open-source tool that provides deterministic, local-first memory and guardrails for AI coding agents via static analysis. Without requiring an API key, it builds a knowledge graph of the codebase, enabling agents to start tasks already oriented to relevant functions, callers, and risks, and to certify changes before they land. It suits large or private codebases and has been shown to reduce agent round-trips by 26% on deep tasks.

Pico is an AI-powered color corrector built on Claude, designed to streamline footage standardization without removing creative control. It automates the conversion from S-Log3 to Rec.709 color space, leaving the artistic grading decisions to the user.

Microsoft maintained its capital expenditure forecast for the year, surprising investors and boosting its stock by 8%. While rivals like Alphabet, Tesla, and Meta increased AI spending, Microsoft's steady capex may signal a slight pullback in capacity expansion due to rising memory costs.

The world's most valuable chip stocks have lost over $1 trillion in market cap since Friday, led by AI-related names. Analysts say the selloff is sentiment-driven rather than fundamental, and AI demand remains strong long-term.

Microsoft is struggling to patch vulnerabilities found by Anthropic's AI model Mythos, which is uncovering bugs faster than the company can fix them, leading to a "mad dash" by engineers.

Microsoft CEO Satya Nadella confirmed during an earnings call that the company is developing a "super app" combining Copilot's chat, coding, and agentic capabilities, set to launch this year for both consumer and commercial use.

Cerebras and AMD announce a partnership to build the world's fastest disaggregated AI inference solution, combining AMD Helios for pre-fill and Cerebras Wafer-Scale Engine for decode, achieving 5x higher tokens per second per watt. The solution targets demanding workloads like agentic coding and real-time voice, with plans to deploy AMD Helios in Cerebras data centers later this year.