Hallucinations
306 episodes
#6506: Why Generative Audio Restoration Hallucinates Highs
Hardware bandwidth limits force modern neural denoisers to hallucinate missing high frequencies from scratch rather than clean real audio.
#6498: Why You Must Read AI Traces Before Writing Evals
Reading just thirty real interaction traces by hand reveals failure modes that generic automated evaluation metrics consistently obscure.
#6497: Why Speech Recognition Confidence Scores Lie
Cross-entropy loss rewards saturation, leaving speech recognition models to assign winning tokens over 0.9 probability even when they are guessing.
#6495: Teaching Speech Models to Choose Their Own Audio
Selecting the most uncertain speech recordings often just surfaces silence, clipping, and abandoned half-sentences.
#6479: Why Research Agent Agreement Is Not Independence
When two distinct language models fail a task, they give the exact same wrong answer sixty percent of the time.
#6473: Why Whispering Breaks Speech Recognition Physics
Whispering completely removes fundamental frequency and pitch, forcing speech recognition models to rely on subtle duration cues.
#6470: Why Computer Science Publishes in Conferences
When NeurIPS tested its review process with duplicate submissions, half the papers accepted by one committee were rejected by another.
#6458: Why Synthetic Fine-Tuning Data Needs Tooling
Rewriting high-quality data synthetically gives a 3.4-point gain, but running the same process on low-quality data yields just 0.5.
#6398: Why Named Entity Linking Fails After the Span Is Found
Zero-shot large language models still lag behind supervised systems on named entity recognition benchmarks by over eleven F1 points.
#6393: Why Model Verbal Tics End Up Driving Script Benchmarks
A pattern search across forty recent episodes found host emotional forecasting in 55% of scripts generated by GPT-6.1 Sol.
#6392: Verifying Generated Arabic Text as a Non-Reader
A study of DALL-E 3 across 2,808 images found its correctness on Arabic letter forms was effectively zero.
#6390: Why Editing Vector Databases Silently Re-Embeds Text
A one-word edit cannot nudge existing vector numbers; the model must re-embed the chunk from scratch or search answers with deleted text.
#6362: Where Mathematicians Actually Work and How AI Fits In
A surge of over 700 automated AI proof manuscripts hit the mathematics community in October, testing how researchers verify claims.
#6319: Tuning Hilbert Back from the Surreal
When prompted simply to be ridiculous, an AI persona began awarding household voting rights to an upstairs staircase landing.
#6305: Why Terminal Fuzzy Finders Forgive Different Mistakes
Fred Damerau found that over eighty percent of spelling errors in retrieval systems were a single insertion, deletion, substitution, or transposition.
#6303: Blind Testing AI Video Models for Architects
A matrix of sixteen prompts and two seeds across three video models produces ninety-six clips to generate, fund, and review.
#6302: Why Fine-Tuning Whisper Won't Stop Duplicated Text
Whisper evaluates transcripts against a gzip compression threshold of 2.4, rejecting repetitive text before retrying at higher temperatures.
#6296: Why Speech Recognizers Cut Off Hesitant Speakers
Between speakers, conversational handover averages 200 milliseconds, but planning pauses inside a turn naturally take seconds.
#6274: Machine Learning for Smells, Whale Codas, and Aftershocks
A graph neural network mapped molecular structures to scents, matching a consensus panel better than the typical individual human nose.
#6240: Gold vs Silver Training Data: What Makes Labels Trustworthy
Human-checked "gold" labels and machine-made "silver" ones aren't as different as the names suggest — and the gap matters.
#6239: Testing Synthetic Data Generators Before You Scale
A benchmark trains student models on each generator's output to find out which model actually writes useful training data.
#6215: Can Anyone Still Spot an AI Voice?
Clones now pass as human. Generic AI voices still don't. And the detectors? They fall apart in the wild.
#6213: Ten Weirdest AI Evals Ever Published
A pelican on a bicycle, a badger for Groundhog Day, and a benchmark where passing is the bad outcome — the strangest AI evals ever run.
#6212: Woolly Sedate Snail and Other Model Failures
Most of Hugging Face's 2.96M models sit unread. But the dust has a personality — snails, pirates, and a matchbox that resigns.