Despite its 31 days, December is a brief month. It’s exhausting for bulletins and happenings aside from workplace events to get consideration. Fighting this development, OpenAI made a sequence of bulletins: their “12 Days of OpenAI.” Not to be upstaged, Google responded with a flurry of bulletins, together with their Gemini 2.0 Flash Thinking mannequin. Models appeared that would use streaming audio and video for each enter and output. But maybe a very powerful announcement was DeepSeek-V3, a really giant mixture-of-experts mannequin (671B parameters) that has efficiency on a par with the opposite high fashions—however value roughly 1/tenth as a lot to coach.
AI
- DeepSeek-V3 is one other LLM to observe. Its efficiency is on a par with Llama 3.1, GPT-4o, and Claude Sonnet. While coaching was not cheap, the value of coaching was estimated to be roughly 10% of the larger fashions.
- Not to be outdone by Google, OpenAI previewed its subsequent fashions: o3 and o3-mini. These are each “reasoning models” which have been skilled to resolve logical issues. They could also be launched in late January; OpenAI is searching for security and safety researchers for testing.
- Not to be outdone by 12 Days of OpenAI, Google has launched a brand new experimental mannequin that has been skilled to resolve logical issues: Gemini 2.0 Flash Thinking. Unlike OpenAI’s GPT fashions that assist reasoning, Flash Thinking reveals its chain of thought explicitly.
- Jeremy Howard and his group have launched ModernBERT, a serious improve to the BERT mannequin they launched six years in the past. It is available in two sizes: 139M and 395M parameters. It’s perfect for retrieval, classification, and entity extraction, and different parts of an information pipeline.
- AWS’s Bedrock service has the power to examine the output of different fashions for hallucinations.
- To be sure they aren’t outdone by 12 Days of OpenAI, Google has introduced Android XR, an working system for prolonged actuality headsets and glasses. Google doesn’t plan to construct their very own {hardware}; they’re partnering with Samsung, Qualcomm, and different producers.
- Also to not be outdone by 12 Days of OpenAI, Anthropic has introduced Clio, a privacy- preserving strategy to discovering out how individuals use their fashions. That data will probably be used to enhance Anthropic’s understanding of questions of safety and to construct extra useful fashions.
- Not to be outdone by 12 Days of OpenAI, Google has introduced Gemini 2.0 Flash, a multimodal mannequin that helps streaming for each enter and output. The announcement additionally showcased Astra, an AI agent for smartphones. Neither is usually out there but.
- OpenAI has launched canvas, a brand new characteristic that mixes programming with writing. Changes to the canvas (code or textual content) instantly change into a part of the context. Python code is executed within the browser utilizing Pyodide (Wasm), moderately than in a container (as with Code Interpreter).
- Stripe has introduced an agent toolkit that permits you to construct funds into agentic workflows. Stripe recommends utilizing the toolkit in take a look at mode till the applying has been totally validated.
- Simon Willison reveals easy methods to run a GPT-4 class mannequin (Llama 3.3 70B) on a fairly well-equipped laptop computer (64GB MacBook Pro M2).
- As a part of their 12 Days of OpenAI sequence, OpenAI lastly launched their video era mannequin, Sora. It’s free to ChatGPT Plus subscribers, although restricted to 50 five-second video clips monthly; a ChatGPT Pro account relaxes lots of the limitations.
- Researchers have proven that superior AI fashions, together with Claude 3 Opus and OpenAI o1, are able to “scheming”: working towards the pursuits of their customers to attain their targets. Scheming contains subverting oversight mechanisms, deliberately delivering subpar outcomes, and even taking steps to stop shutdown or substitute. Hello, HAL?
- Roaming RAG is a brand new approach for retrieval augmented era that finds related content material by looking by way of headings to navigate paperwork—like a human may. It requires well-structured paperwork. A surprisingly easy thought, actually.
- Google has introduced PaliGemma 2, a brand new model of its Gemma fashions that includes imaginative and prescient.
- GPT-4-o1-preview isn’t any extra; the preview is now the true factor, OpenAI o1. In addition to superior reasoning expertise, the manufacturing launch claims to be sooner and to ship extra constant outcomes.
- A bunch of AI brokers in Minecraft behaved surprisingly like people—even creating jobs and religions. Is this a method to mannequin how human teams collaborate?
- One factor the AI trade wants desperately (except for extra energy) is higher benchmarks. Current benchmarks are closed, simply gamed (that’s what AI does), and unreproducible, they usually could not take a look at something significant. Better Bench is a framework for assessing benchmark high quality.
- Palmyra Creative, a brand new language mannequin from Writer, guarantees the power to develop “style” so that each one AI-generated output received’t sound boringly the identical.
- During coaching AI picks up biases from human information. When people work together with the AI, there’s a suggestions loop that amplifies these biases.
Programming
- Unicon could by no means change into one of many high 20 (or high 100) programming languages, however it’s a descendant of Icon, which was all the time my favourite language for string processing.
- What do CAPTCHAs imply when LLM-equipped bots can efficiently full duties set for people?
- egui, along with eframe, is a GUI library and framework for Rust. It’s transportable and runs natively (on macOS, Windows, Linux, and Android), on the internet (utilizing Wasm), and in lots of recreation engines.
- For the archivist in us: The Manx challenge isn’t about an island within the Irish Sea or about cats. It’s a catalog of manuals for previous computer systems.
- Cerbrec is a graphical Python framework for deep studying. It’s aimed toward Python programmers who don’t have ample experience to construct functions with PyTorch or different AI libraries.
- GitHub has introduced free entry to GitHub Copilot for all present and new customers. Free entry provides you 2,000 code completions and 50 chat messages monthly. They’ve additionally added the power to make use of Claude 3.5 Sonnet along with GPT-4o.
- Devin, the AI assisted coding device that claims to assist software program improvement from starting to finish, together with design and debugging, has reached normal availability.
- JSON5, also referred to as “JSON for humans,” is a variant of JSON that has been designed for human readability in order that it may be written and maintained by hand—for instance, in configuration information.
- AWS has introduced two important new companies: Aurora DSQL, which is a distributed SQL database, and S3 Tables, which helps information lakehouses by way of Apache Iceberg.
- AutoFlow is an open supply device for making a information graph. It’s primarily based on TiDB (a vector database), LlamaIndex, and DSPy.
Security
- Portspoof is a safety device that causes all 65,535 TCP ports to look open for legitimate companies. It emulates a legitimate service on each port. It makes it troublesome for an attacker to find out which ports are literally open with out probing every port.
- Let’s Encrypt, which points the certificates that web sites (and different functions) use to show their identities, has introduced short-lived certificates that expire after six days. Short-lived certificates enhance safety by minimizing publicity if a non-public key’s compromised.
- Because of the continued presence of attackers inside telecommunications networks, the US FBI and CISA have really helpful using encrypted communications protocols. (Though they nonetheless need backdoors into encryption methods, which might make them weak to assault.)
- A new phishing assault makes use of corrupted Word paperwork to bypass safety checks. While the paperwork are corrupt, Word is ready to recuperate them.
- LLM Flowbreaking is a brand new class of assault towards language fashions that forestall guardrails from stopping objectionable output from reaching the person. These assaults reap the benefits of race circumstances within the utility’s interplay with customers.
- Bootkitty is a UEFI bootkit that targets safe boot on Ubuntu methods. It seems to have been developed by cybersecurity college students in Korea, then leaked (presumably unintentionally). It hasn’t but been discovered within the wild, however when it’s, it will likely be a harmful menace.
- DEF CON has began a challenge to enhance cybersecurity for water infrastructure within the US. They’re beginning with six water corporations serving rural communities.
Quantum Computing
- Google has constructed a quantum computing chip through which an error-corrected logical qubit can stay secure for an hour. It passes the “below threshold”: the error charge decreases as bodily qubits are added for error correction. The chip was in-built Google’s new fabrication facility.
Web
- Google is including “store reviews” to Chrome. Reviews are AI-generated summaries of experiences from well-known sources that report scams and different points.
- Here’s a how-to on constructing streaming textual content person interfaces on the internet. Streaming textual content is nearly a necessity for constructing AI-driven chatbots.
Biology
- Yes, we are able to have digital style. A analysis group has developed a lollipop interface so that folks can expertise style in digital worlds.
