Jun 12, 2025

Daily Updates

Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents

• The Darwin Godel Machine (DGM) represents a breakthrough in self-improving AI, enabling systems to iteratively modify their own code and validate changes through empirical benchmarks, thus enhancing their coding capabilities significantly.

Chatterbox TTS

• Chatterbox is an open-source TTS model from Resemble AI, outperforming closed-source systems like ElevenLabs in user evaluations, and features emotion exaggeration control for enhanced expressiveness.

Seedance 1.0

• Seedance 1.0 revolutionizes video generation by enabling multi-shot creation from text and images, achieving 1080p resolution with cinematic quality and smooth motion.

EchoLeak – 0-Click AI Vulnerability Enabling Data Exfiltration from 365 Copilot

• Aim Labs has uncovered "EchoLeak," a zero-click AI vulnerability in Microsoft 365 Copilot that allows attackers to exfiltrate sensitive data without user interaction, leveraging a novel technique called "LLM Scope Violation." This vulnerability exploits design flaws in RAG-based chatbots, highlighting significant risks in AI applications.

Maximizing Battery Storage Profits via High-Frequency Intraday Trading

• Automated high-frequency trading strategies for battery energy storage systems can significantly enhance revenue by reacting swiftly to market changes, achieving a 58% increase in profits compared to hourly re-optimizations.

Fine-tuning LLMs is a waste of time

• Fine-tuning LLMs is not knowledge injection; it risks overwriting existing, valuable information within the model, leading to unexpected and problematic outcomes.

[R] Semantic Drift in LLMs Is 6.6x Worse Than Factual Degradation Over 10 Recursive Generations

• Semantic drift in LLMs is 6.63 times worse than factual degradation, with a 42.5% drop in Purpose Fidelity over 10 recursive generations, indicating a significant loss of intended meaning despite factual accuracy remaining largely intact.

Institutional Books: A 242B token dataset from Harvard Library's collections

• Institutional Books 1.0 is a refined dataset of 242 billion tokens derived from 983,004 public domain volumes digitized by Harvard Library, enhancing the quality and usability of training data for large language models (LLMs).

First thoughts on o3 pro

• OpenAI's o3 pricing was slashed by 80%, launching o3-pro at $20/$80, which reportedly achieves a 64% win rate against its predecessor, indicating a significant leap in performance for task-specific models.

[R] FlashDMoE: Fast Distributed MoE in a single Kernel

• FlashDMoE achieves unprecedented efficiency by fusing the Distributed MoE forward pass into a single kernel, resulting in up to 9x higher GPU utilization and 6x lower latency.

2025 State of AI Code Quality

• AI coding quality in 2025 hinges on developer confidence rather than mere code generation, with 65% of developers citing a lack of relevant context as a major issue affecting trust in AI outputs.

[R] ABBA: Highly Expressive Hadamard Product Adaptation for Large Language Models

• ABBA introduces a novel architecture for Parameter-Efficient Fine-Tuning (PEFT), outperforming LoRA and its variants by reparameterizing updates as a Hadamard product of two independently learned low-rank matrices, enhancing expressivity and performance.

[D] Image generation using latent space learned from similar data

• Image generation can be enhanced by leveraging latent space differences between cell phases, potentially allowing for the creation of phase 2 cell images from phase 1 data using a VAE (Variational Autoencoder) approach.

[P] SWE-rebench Major Update: Tool Usage, Claude Sonnet 3.5/4, OpenAI o3 and May Data

• SWE-rebench has introduced Tool Usage Support, allowing agents to engage with environments through both text and tools, enhancing the benchmarking process for software engineering LLMs.

Revisit What You See: Disclose Language Prior in Vision Tokens for Efficient Guided Decoding of LVLMs

• ReVisiT introduces a novel decoding method that effectively utilizes vision tokens to enhance text generation in Large Vision-Language Models (LVLMs), improving visual grounding without extensive retraining or complex procedures.