Overtrained Language Models Are Harder to Fine-Tune Springer et al.: https://
arxiv.org/abs/2503.19206 #ArtificialIntelligence #Deeplearning #MachineLearning
RESEARCH
-

Overtrained Language Models Are Harder to Fine-Tune
By
–
-
AI Mirrors Society’s Biases: LLMs Struggle to Keep Pace
By
–
AI acts as a mirror. It reflects our society's pre-conceived notions and biases right back at us. Societies keep evolving and LLMs will not be able to keep pace. So, here’s what we need to do…
-
AI Models Implement Bias Mitigation Strategies for Improvement
By
–
Not all AI models suffer from these biases. Many of them have implemented various mitigation strategies to address such issues. These include ‘Fine-tuning’, ‘Reinforcement Learning from human feedback’, ‘Constitutional AI’, and ‘debiasing’. LLMs are steadily improving. As they
-

Addressing AI Image Generation Bias in Diverse Demographics
By
–
The bad news is AI has biases. The good news is we can fix them. I asked a specific model to generate an image of senior Asian executives sitting around a round table. I won't name the model, as the idea is to focus on the issue rather than criticize any company. The model
-

Google Leads on Multiple Fronts in AI Metrics
By
–
Google is now leading on multiple fronts on AI metrics.
-

OpenAI Launches GPT-4.1 Using Scale AI Benchmark
By
–
Congrats to @OpenAI on the successful launch of their GPT-4.1 model 🙂 Great to see their team utilize the @scale_AI MultiChallenge benchmark to measure multi-turn instruction following
-
Whisper’s Hallucinations: Understanding ASR Limitations
By
–
We dive into the well defined task of automatic speech recognition (ASR), and describe why OpenAI’s Whisper, which has been integrated into ChatGPT, makes stuff up, or “hallucinates” as it’s called in the industry (bad nomenclature).
-
One-Model-Everything Approach Reduces ASR System Reliability
By
–
Their (and Muskrat and others’) quest to build one-model-for-everything has resulted in less reliable systems than what we’ve had before, even in the well defined task of ASR. As in, historically “hallucinations” weren’t problems in ASR systems!
-
Why AI Speech Recognition Got Worse Despite Hype
By
–
In Scientific American, @asmelashteka & I ask: "In the age of unbridled AI hype, with the likes of Elon Musk claiming to build a “maximally truth-seeking AI,” how did we come to have less reliable speech recognition systems than we did before?"
-

GPT-4.1 Launch: New Model for Developer Excellence
By
–
Welcome to the GPT-4.1 family! Trained with one goal: being truly great for developers. • Best-in-class at coding tasks and instruction following
• 1M token context + better comprehension
• Pushing performance across the latency curve What are your first impressions so far?