These sources apply machine learning to a particular job, from chemistry to images. This page lists eleven sources, oldest first where dated.
Template Matching Advances and Applications in Image Analysis
Four authors from one university survey basic template matching measures and invariance methods. The paper then summarizes applications from licence plates to brain imaging. It reports others' results and runs no new experiments of its own in the text.
Dishwashing Robot for Restaurants
The post is a few paragraphs long and gives no figures. It lists the parts of the Dishcraft system and how a dish moves through it. The page also carries affiliate ads for robot toys.
How we made our optical character recognition (OCR) code more accurate
The post comes from Pieces, a company that sells a developer app. It describes pre-processing steps and an indentation fix added around Tesseract. The evidence is a described test method, with no scores shown.
A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists
A peer-reviewed paper that builds a chemistry benchmark. It tests leading language models against volunteer chemists. It reports the best models beat the best human overall. They also fail some basic tasks and misjudge their own accuracy. The evidence is benchmark scores, a small human survey and public code.
Nano Banana can be prompt engineered for nuanced AI image generation
A Hacker News thread on Max Woolf's 26 minute post about prompting Nano Banana. That is Google's Gemini 2.5 Flash Image model. The post gives prompts, generated images and a Python wrapper. Commenters add their own tests, complaints, workflows and a long argument about AI and art.
Post by @emollick.bsky.social on a Claude Code adventure game
A short social post with a screenshot and a link to a playable game. The poster says Claude Code made the game from one prompt. Two more prompts tested and deployed it. The evidence is the poster's report and a link that was not opened here.
Reimagining the mouse pointer for the AI era
A Google DeepMind blog post that describes an experimental pointer powered by Gemini. It sets out four interaction principles and says the ideas are entering Chrome and Googlebook. The evidence is demo videos and product announcements, with no user study.
AI as a Tool for the Mental Load
A first-person essay by Brittany Ellich about using Claude for trips and meals. It argues the main benefit is relief from tracking tasks, not productivity. The evidence is the author's own routine and one anecdote about a parent.
Google DeepMind Maps 9 Billion Possible DNA Variants
A three minute news article on the AlphaGenome Atlas, announced on 8 September. It explains what the Atlas holds, how it was computed, and its stated limits. Sources are two DeepMind staff and one outside genomicist.
Show HN: Happy Hangul Day! An RNN for Generating Korean Handwriting Strokes
A Show HN thread for hangul.ink, a model that writes Hangul as animated pen strokes. The author's blog post covers the data and design. Comments report bugs and ideas. The evidence is a working demo, an open dataset and informal user testing.
ai-music-video (repository skill-generative-video-editing)
A small repository by faz.ms with a README and a SKILL.md. They instruct a coding agent how to make a music video. It covers timing, keyframes, video clips, rough cuts and compositing. The evidence is the instructions themselves, with no example output read.
Glossary
- Autoregressive
- Generating output one piece at a time, each depending on the earlier pieces.
- Base pair
- One rung of the DNA ladder, a matched pair of letters.
- Benchmark
- A fixed set of questions used to compare how well different systems perform.
- Bicubic upsampling
- A standard way to enlarge a small image by smooth interpolation.
- Calibration
- How well a stated confidence matches how often the answer is right.
- Claude
- An AI assistant made by Anthropic.
- Claude Code
- An AI coding tool from Anthropic that works from a terminal.
- Compositing
- Combining separate images or clips into one frame.
- Computer vision
- Software that interprets what cameras see.
- Deformable template
- A template that can bend or stretch to fit varying shapes.
- Dice coefficient
- A score of overlap between two regions, from 0 to 1.
- EGA
- A 1980s graphics standard with 16 colours, used by early adventure games.
- Enhancer
- A stretch of DNA that boosts the activity of a gene, sometimes from far away.
- Entity
- A recognised item such as a place, date or object that software can act on.
- Gemini
- Google's family of AI models.
- Google AI Studio
- A Google web tool for trying and building with its AI models.
- Google Flow
- Google's tool for generating images and video from prompts.
- Googlebook
- A laptop experience the post names, with Magic Pointer arriving soon.
- Hangul
- The Korean alphabet, created in 1446.
- Inpainting mask
- A marked region that tells an image tool where edits are allowed.
- Instacart
- A grocery delivery and pickup service.
- Invariance
- Giving the same result despite a change such as scale or rotation.
- Keyframe
- A still image that fixes how a shot looks at its start or end.
- Levenshtein distance
- A count of the edits needed to turn one text into another.
- LoRA
- A small add-on that teaches an image model a specific subject or style.
- LSTM
- A type of neural network that reads sequences, here to predict characters.
- Mental load
- The unseen work of noticing, remembering and planning what a household needs.
- Mixture density network
- An output layer that predicts several possible next values and samples one.
- Model distillation
- Training a smaller, faster model to copy a larger one.
- Nano Banana
- The popular name for Google's Gemini 2.5 Flash Image model.
- NMR
- Nuclear magnetic resonance, a lab method that reveals a molecule's structure.
- Noncoding DNA
- DNA that does not code for a protein but may regulate genes.
- Normalized cross-correlation
- A similarity score between -1 and 1 that ignores overall brightness changes.
- OCR
- Software that turns images of text into editable text.
- Petabyte
- A million gigabytes.
- Prompt engineering
- Writing and refining the text given to an AI model to get better output.
- Retrieval augmented generation
- Giving a model relevant documents to look up before it answers.
- RNN
- A recurrent neural network, a model that loops while keeping a memory of earlier steps.
- Rough cut
- A first assembly of clips edited against the music.
- Show HN
- A Hacker News post category for projects the poster built.
- Skill
- A folder of instructions and scripts that teaches Claude Code a specific task.
- Smartpen
- A pen that records its position and pressure as you write.
- SMILES
- A text notation that writes a molecule's structure as a string.
- Style transfer
- Redrawing an existing image in another artistic style.
- Subagent
- A helper AI process that the main AI can call for a specific job.
- Template matching
- Finding where a small reference image appears inside a larger image.
- Tesseract
- A free OCR engine that the post uses as its base.
- Text parser
- Game code that reads typed commands such as open door.
- Vacuum gripper
- A robot hand that holds objects by suction.
- Variant
- A place where a person's DNA differs from the reference sequence.
- Vibe-coding
- Building software by describing it to an AI model that writes the code.