Could there be an LLM built into everything programmable? It seems like it could be. llama.ttf is a plain old font that looks like Open Sans. You can download it and install it. I don’t know if this is a massive vulnerability or a way to embed AI in almost any text-based application. It’s probably both.
It may be a coincidence, but several new programming languages have appeared in the past month or two. Why? Probably not because of a backlash against automated programming (LLMs can’t be trained in languages that don’t have a lot of open source code).
Learn faster. Dig deeper. See further.
AI
- This is crazy. llama.ttf is a font with a (small) large language model built in. The font itself can do automatic text generation. It relies on the HarfBuzz font shaping engine using Wasm. It works in Gimp. It can also work in Chrome and Firefox. Everything runs locally.
- Researchers found that talking to AI can help dispel beliefs about conspiracy theories. (The paper notes that the results are preliminary and have not yet been peer-reviewed. All data from the study will be published once it is complete.)
- Podman AI Lab is a great way for Linux users to experiment with running AI locally. Unsurprisingly, Podman is container-based, making it easy to move models from desktop to production.
- Researchers developed a BERT-based model to detect malicious LLM prompts that instruct LLM to create phishing websites or emails.
- Anthropic has released a new version of their medium-sized language model, Claude 3.5 Sonnet. They claim similar performance to GPT-4o and Gemini 1.5. A new Artifacts window lets you interact with Claude output.
- What if non-player characters in a game could talk to you and each other? What if the game’s structure was not scripted in advance, but unfolded dynamically as you played? That’s what generative AI can bring to games.
- AI systems have been trained to count flowers, an ability that has important applications in agriculture: counting flowers can help predict crop yields.
- Anthropic has provided documentation on how to build evaluation and test cases for Claude.
- Meta has developed AudioSeal, a system that watermarks AI-generated audio content. It has 90% to 100% reliability and can identify small clips embedded in larger files. AudioSeal is free and available on GitHub, but Meta itself does not use it yet.
- Cohere’s latest model, Command R, can respond to prompts using multiple external tools. It allows you to create a plan to solve a problem, access external tools using an API (described in the JSON document), and use multi-step tools to generate results.
- Can a large language model write comedy? Obviously not. The ideas are unoriginal, the model has trouble with inconsistencies and surprises, and has no sense of the long arc of conversation.
- Luma’s Dream Machine is an impressive generative AI tool that creates short videos from text prompts.
- NVIDIA has released a series of models for generating synthetic training data. The use of the models is governed by a new Open Model License Agreement, which states that the models can be used commercially, derivative models can be freely distributed, and NVIDIA does not claim ownership of the output of the models.
- Contrary to vendor claims, AI detectors with very high error rates are being used to unfairly block professional writers.
- A novel strategy for chunking content in RAG is to compute embeddings for each sentence using SBERT (sentence-BERT), and then use the embeddings to detect topic changes.
- When building AI applications with RAG, consider using a graph database (GraphRAG) instead of a vector database. Graphs are harder to work with, but they are better at providing the right context to answer complex queries.
- Google’s experimental Illuminate turns academic papers into interview-like discussions that summarize the key points. It’s currently in private beta.
- Anthropic’s essay on Claude’s character is a thoughtful piece on how to give Claude three character traits and how character traits affect the way Claude interacts with the world.
- Honeycomb’s VP of engineering said AI will be more helpful for developing front-end code than back-end code, and that chat is often an inefficient user interface.
- OpenAI and Anthropic have published a study that extracts features from the GPT and Claude models, which represents an important step toward interpretability, or understanding why a model gives a particular response to a prompt.
- What is your dog trying to say? A research project uses AI to decode dog vocalizations.
- Let’s take a look at the Transformer-based language model that Apple is integrating into macOS. It’s very small (34M parameters, apparently) and appears to be based on GPT-2. It’s clearly designed to run on devices like phones without draining the battery.
- Codestral is a new language model from Mistral that specializes in code generation. It has been trained on over 80 languages and claims to outperform other language models.
- Language models can generate database schemas, but schemas rarely remain unchanged forever. How well can language models modify schemas?
- Can AI help researchers digitize smells? Digitization has been effective for sight and hearing. Smell is lagging far behind.
- Could AI improve privacy? Researchers have developed an application that removes sensitive parts of an image and replaces them with something similar but different. It’s a kind of automated Photoshop.
- In the context of RAG, using knowledge graphs makes intuitive sense. When comparing “GraphRAG” to “traditional RAG”, GraphRag is superior because the graph encodes relationships between data.
programming
- Gleam is a new type-safe functional language. It compiles to Erlang and JavaScript. Its goal is to enable developers to build large-scale concurrent systems.
- OpenTofu, an open source fork of Terraform, now adds state encryption, a feature that has been requested by Terraform for nearly a decade.
- Starlark is a small Python-like language that describes how to build code. Starlark’s design principles include determinism and safety. Multiple executions of a program will produce the same results, and it cannot access the network, file system, or system clock.
- One of the co-creators of WebAssembly has developed a new programming language called Virgil. It is designed to be a systems programming language, but is not intended to compete with Zig (whatever that means).
- The James Webb Space Telescope is a step forward in digital twin technology. A digital model of the telescope was used to coordinate its self-assembly in space. The twin continues to monitor the telescope and predict the effects of software updates.
- Mesop is a very lightweight Python framework for rapid web application development. It’s good for demos, but probably not good for production. It’s built by Google, but it’s not supported.
- AI Gateway is similar to API Gateway, but is specialized for accessing APIs of language models. It facilitates monitoring, selecting models to serve requests, caching, security, and many other functions.
- Julia Evans (@b0rk)’s latest magazine, How Git WorksThis is a must-read for anyone who doesn’t understand Git. Which is pretty much everyone.
- The language developers have been busy. The Grain language is designed to compile to WebAssembly. It is inspired by OCaml and pattern matching is a key feature.
- GoFr is a framework for building microservices in Go. It integrates tools for observability and supports using a variety of data sources.
- The approach of finding infrastructure in code argues that the infrastructure requirements of a system can be inferred from the code itself, and specifications for the required infrastructure can be automatically generated.
security
- It’s unclear what’s happening, but it appears there was a supply chain attack on CDN service Polyfill.io. Polyfill claims it was defamed and moved to a new domain name. Security experts advise caution.
- A new serious vulnerability in Progress Software’s MoveIT product is being actively exploited. Attackers can impersonate legitimate SFTP users to copy, delete, or create files. Attackers can also obtain encrypted hashes of user passwords.
- The US will ban the sale of Kaspersky’s antivirus software starting in July.
- Linux malware controlled by emoji sent from Discord? That’s creative. Spreading via phishing, the malware uses emoji to send commands. Emojis seem to evade security software that expects commands to be text strings.
- A new type of phishing uses the Windows Search protocol to download malware. Another new type of phishing uses Progressive Web Applications (PWAs) to steal credentials.
- After widespread criticism of the AI-assisted Recall feature in Windows 11, Microsoft announced that it would be disabled by default. The release of Recall has been delayed. It will be available first to the Windows Insider Program, which is basically a closed beta.
- As AI becomes more important, Hugging Face has become a target for threat actors. Most recently, attackers stole authorization tokens from Spaces.
knitting
- SimCity WebAssembly in the browser? Impressive.
- UIX is a new full-stack web development framework. It is part of the larger Unyt project, which has the ambitious goal of building the next-generation decentralized internet.
- BBC Ventures invests in 3D video streaming platform
- Hotwire is an approach to building web applications (especially single page applications, or SPAs) that encapsulate complex data by sending HTML rather than JSON. The goal is to minimize the need for JavaScript.
- Cobalt is a very lightweight platform for running apps built with HTML5, CSS, and JavaScript. The creator, Google, cites YouTube as a typical application, but we think it could be used for more interesting things.
- Adrian Holovaty’s Soundslice integrates music scanning. Upload a PDF or photo of your score and Soundslice analyzes it. You can then edit, rearrange, and play the score.
hardware
- Open standards for connectors? That’s important if you want very high performance without having to license proprietary designs.
- Energy Meter is a tool that measures the power consumed by Intel CPUs. It can be useful in today’s power-hungry AI applications, but it does not take GPUs into account. It runs on Linux, but most servers in the world are Linux.
- Mitsubishi has developed a robot that can solve the Rubik’s Cube. 0.305 seconds. Their best time is 0.204 seconds, which according to Guinness was not achieved under ideal conditions.
- On Lake Michigan, rescue robots are being used to help people who have fallen into the water stay afloat until rescuers arrive.
biology
- Precision Neuroscience has successfully implanted an array of 4,096 tiny electrodes into the human brain, achieving higher resolution of neural activity than previous implants, including Neuralink.
