Anthropic announcement of Claude 3.7 Sonet, regardless of this, it seemed that the breakthrough pace of the main announcement AI slows down to February. This gave us the time to look at some other topics. Two important programming posts have appeared: Salvatore Sanfilippo is “We Destruction of Software” and Rob Pike’s Slide Deck “On Bloat”. They are surprisingly similar. Neith mentions AI; Both are wondering why our hardware is accelerating and faster, but our applications are not. We also noticed the return of Pebble, the first intelligent watches and table lamps controlled by AIT from Apple Research, which looks as if it came from the Pixar logo. Entertainment, maybe not to look for an Apple store.
Artificial intelligence
- Anthropic was released by Claude 3.7 Sonet, the first model of justification of the company. Has a “hybrid model”; You can tell where you want to enable its ability to think. You can also control its idea of a “budget” by limiting the number of tokens that generates for the reasoning process.
- Arena computer agent is a platform for testing agent Crowdsourced. It allows you to run the agent using two different AI models, watching what the agent is doing and evaluating the results. The results are summarized, we have a ranking; Right now Claude 3.5 Sonet is up.
- Google is developing “ko-scientists”, assuming scientists to investigate. Hypotheses are based on goals, ideas and past research by a scientist. The company is looking for researchers who help with testing.
- Github upgraded the fashion agent for Copilot. It will now ber on the COGGY CODE unit, which brings the correct results, and can add new partial tasks to the original if they need to achieve the user’s goal.
- Open-R1 is a new project that is interested in creating a fully open reproduction Deeepseek R1. In addition to code and weight, this project releases all tools and synthetic data used to train the model.
- Moshi is a new conversational language model (Speech-Green) that constantly listens and can handle interjections like “uh huh” without being confused.
- Codename Goose is a new open source framework for developing AI agent applications. The context protocol protocol for communication with systems that have data and can discover new data sources while running.
- The University of Surrey will build a language model for sign language. One focus will be the translation between the spoken language and sign language. The aim is to ensure that the deaf community does not remain explosion AOLS.
- Galileo is an agent set of tools for detection when hallucinations of the AI model. It is particularly important for agency systems where one agent’s error leads to bad behavior by other mines.
- A group of scientists released S1, a 32B thinking model with almost the most modern performance. S1 costs only $ 6 for training. A very small set of training data (only 1,000 sampling samples) proved sufficiently when the model was forced to take more time to justify.
- Some scientists have published How to scalpA book on how a scale of expanding models of large languages. The book is obviously internal documentation from Google Deepmind.
- Openai released O3-Minini, a small and cost-effective language model based on its (still non-released) model O3.
- Anthropic deployed its constitutional classification for public contradictory testing. The classification is a system that protects Claude models from prison escape and tries to get Claude to answer questions that are not allowed. The first results look very good.
- The lesson to learn from Deepseek R1 is that due to the good endowment model is less difficult than he thought that the development of a model of reasoning. In the coming months, expect many open alternatives.
- Openai introduced Deeperearch, an application based on its O3 model, which claims that the ability to synthesize great love and research tasks of multi -top research.
- Sam Altman acknowledged that OpenIi was the “bad side of history”, if Aijit also said that the socket solution was not a high priority.
- Alibaba launched Qwen2.5-Max, another large language model with power at the same level as the GPT-4 and Claude 3.5 Sonet. Can be highlighted via Qwen Chat or Alibaba’s Cloud.
- Transform Lab is a tool for experience, training, fine -tuning and programming LLM models locally. It is still installed, but it looks like Ollama on steroids.
- Smolgpt is “At least Pytorch Implementation for training your own small LLM from scratch.”
- Yes, Microsoft complains that Deecseek used synthetic training to generate synthetic training data. These objections did not prevent Deepseek Azure from creating Azure.
- Two composers worked with Google’s Gemini Twin paradoxWork for classic symphony orchestra.
- Alibaba released two “checkpoints”, Qwen2.5-7b-Instruct-1M and Qwen2.5-14b-Instruct-1M. These models have a wide 1m-tet context windows. Alibaba also has an open-source inference frame that claims to be three to seven times faster.
- Tinyzero reproduces Deepseek’s zero R1, a model of thinking with 3B parameters. Tinyzero training is worth $ 30. You could download Tinyzero, but you can also create your own at less event costs. Do we need expensive models?
Programming
- Tanagram promises a set of tools to help developers understand and work with complex codes. So far there are only demos, but it sounds interests.
- Harper Reed describes his workflow for programming with AI. The development of the workflow is essential for efficient use and Harper has given the most expensive description we have.
- Like Linux, Ruby On Rails can run in the browser. This hack worms of the web soucre.
- Booting Linux inside PDF in chrome. Implementation of PDF JavaScript support; C can be compiled into the subset of JavaScript (asm.js), which means that the RISC-V emulator can be compiled to JavaScript and run in PDF in the browser that Lunux. Amazing hack.
- OCR4ALL provides software for recognition without an optical feature and open source. You should need it.
- Why doesn’t software run fast than 20 or 30 years ago, despite much faster computers? Rob Pike has several thoughts to control flatulence.
- As the name suggests, the records of architectural decision -making (ADRS) capture the decision on software architecture and the reason for the decision. All too often this information is not captured. It is likely to become more important at the time of software development with the help of Ai-Aistised.
- Jank is a new programming language of the general purpose. It is a dialect Clojure that has integrated thoughts from many other languages, included C ++ and Rust and is built on the top of LLVM.
- Here is a set of patterns for creating real -time features into applications.
- Post Salvatore “Antires” Sanfilippo, “Destroys Software”, is a necessity. .
- The script is a Library of Go, which allows programming similar to Shell in Go. It also has GO features that are similar to GRP, FIND, Head, Tail and other common shell commands.
Security
- The electioners who are consistent with Russia focus on signal, secure application for sending messages, with phishing attacks that connect users’ accounts to hostile devices. One group sends QR codes that look legitimate but refer to the device under their control; Another one is issued as an application used by Ukraine by the army. The best protection is to update the latest version of Signal.
- Two new injuries were found in OpenSsh. One exhibitions open servers to attack the middle; The second can lead to attacks on the rejection of the service. An update was issued; Install it.
- Darkmind is a new attack against language thinking models. It is possible to create custom applications (such as GPT) with “hidden triggers” that adjust the reasoning process.
- The new type of supplier chain attack involves obtaining abandoned AWS S3 buckets that still hold libraries that are often downloaded. The new owner can put malware into libraries; The original owner who left the bucket cannot repair the damaged libraries.
- Security blocks AI adoption, especially in cumbersome regulated industries. This is understandable; Many questions we ask on secure systems cannot be adequately answered.
- Published by AI Red Microsoft AI team Lessons from the red team of 100 generative products AI. It is a necessary reading for anyone interests in building a safe AI system.
- AI is used to present fake requirements for functions and reports on open source projects. Many of them can be unintentional, but the appearance of the cause, it generates problems for software stays.
- Linux has a number of tools for detection of rootkits and other malware. Chkrootkit and LMD (Linux Malware Detect) are worth your attention.
- Time Bandit is a new escape from prison for GPT models. The attack causes the model to lose an overview of the past, present and the future. Basically, you ask GPT how someone would have done something in the past that can only be at present. It is not clear whether this attack works on other models.
- When the price of bitcoins rises, the despence frequency of cryptojacking: kidnapping computers to create a crypto mining of botnets. It is said that for every dollar of crypto mining, the victim of $ 53 in the cloud costs causes a victim.
- In the wild, a new backdoor to VPN has been discovered, allowing attackers access to corporate networks. These rear doors remain a sleeping unit, which are lowered by a specially built “magic packet”, which makes it difficult to detect.
Site
- Since more people ask AI to recommend products, traders will have to optimize the perception of the product according to language models. Des LLMO replaces SEO? Optimization for LLM can be another generation of SEO.
- This article tells you how to choose Gemini in Gmail and other Google Workspace applications. It is possible to selectively disable gemini. It requires you to have access to the administration console.
- JavaScript’s temproal object begins to appear in browsers! Time is a replacement for insufficient data object. It allows programs to work efficient with data and times.
- Marginalia is a search engine with an open source code that prefers non -material centers.
Quantum
- Microsoft created topological Qubit on a new quantum chip. While its chip currently has only 8 qubits, Microsoft claims that it can scal millions of Qubits. Placing these many QUBITS on the chip would be a long way to solve the problem of moving quantum data between chips.
- The Canadian startup Xanadu created a quantum computer using photonics. He currently has 12 Qubits, but the company believes that it can scalance for larger systems.
Robotics
Gadget
- Pebble returns? Remember the crowdfunded Pebble Smartwatch, which was available long before the Watch Apple? It will come back – maybe. And it will be hackable.
- Something we all need: Apple’s engineering team developed a desktop lamp controlled AIT. It is not available at the Apple store in your area.
Learn faster. Kick deeper. See on.