Radar trends to see: January 2025 – O’Reilly

Despite its 31 days, December is a short month. It is difficult for ads and events that are not office parties to get attention. Fighting for this trend, Openai made a series of ads: his “12 days of Openai”. In order not to be eclipsed, Google responded with a burst of ads, including its Flash Gemini 2.0 thought model. Models appeared that could use audio and transmission video for input and output. But perhaps the most important announcement was Deepseek-V3, a very large mixture model of experts (671b parameters) that has a performance with the other main models, but costs approximately 1/10 to train.

AI

• Deepseek-V3 It is another Llm to look. His performance is on par with flame 3.1, GPT-4o and Claude Sonnet. While training was not economical, the Training cost It was estimated at approximately 10% of the largest models.
• It should not be overcome by Google, Openai preview Its next models: O3 and O3-mini. Both are “models of reasoning” that have been trained to solve logical problems. They can be released at the end of January; Operai is looking Security researchers For tests.
• To not be left behind for 12 days of OpenAi, Google has launched a new experimental model that has been trained to solve logical problems: Gemini 2.0 flash thinking. Unlike OpenAi GPT models that admit reasoning, flash Thinking shows its chain of thought explicitly.
• Jeremy Howard and his team have Modernbert releasedAn important improvement For the Bert model, they launched six years ago. It comes in two sizes: 139m and 395m parameters. It is ideal for the recovery, classification and extraction of entities, and other components of a data pipe.
• AWS’s mother rock service has the ability to Verify the output of other models For hallucinations.
• To ensure that they are not surpassed for 12 days of OpenAi, Google has Android XR announcedAn operating system for headphones and extended reality glasses. Google does not plan to build its own hardware; They are associated with Samsung, Qualcomm and other manufacturers.
• Nor should the 12 days of Openai be exceeded, Anthrope has announced ClioA privacy conservation approach to discover How people use their models. This information will be used to improve the understanding of Anthrope’s security problems and to build more useful models.
• The 12 days of OpenAi should not be overcome, Google has announced Gemini 2.0 Flash, a multimodal model that supports the input and output transmission. The announcement is also shown StarAn AI agent for smartphones. Neither of them is generally available.
• Operai has launched canvasA new feature that combines programming with writing. Changes on the canvas (code or text) immediately become part of the context. The Python code is executed in the browser using pyoduro (Wasm), instead of a container (as with the code interpreter).
• Stripe has announced a Agent tools set That allows you to develop payments in agent workflows. Stripe recommends using the tool kit in test mode until the application has been thoroughly validated.
• Simon Willison shows How to execute a GPT-4 class model (call 3.3 70b) on a reasonably well equipped laptop (64GB MacBook Pro M2).
• As part of his 12 days of Operai series, Operai finally launched his video generation model, Sora. It is free for chatgpt plus subscribers, although limited to 50 five -second video clips per month; A Chatgpt Pro account relaxes many of the limitations.
• Researchers have shown that advanced AI models, including Claude 3 Opus and Openai O1, are capable of “intriguing“: Work against the interests of its users to achieve its objectives. The scheme includes subverting the supervision mechanisms, intentionally delivery of lower results and even taking measures to avoid off or replacement. Hello, Hal?
• Roaming rag It is a new technique for the generation of augmented recovery that finds relevant content when searching through headers to navigate documents, such as a human. Require well structured documents. A surprisingly simple, really.
• Google has announced Paligemma 2A new version of its Gemma models that incorporates vision.
• GPT-4-O1-Preview does not exist; Previous view is now the real thing, OPENAI O1. In addition to advanced reasoning skills, production launch claims to be faster and offer more consistent results.
• A group of AI agents in Minecraft He behaved surprisingly as humans–In the development of jobs and religions. Is this a way of modeling how human groups collaborate?
• One thing that the AI ​​industry needs desperately (apart from more power) is best reference points. The current reference points are closed, play easily (that is what AI does) and is not reproducible, and they may not try anything significant. Best bank It is a framework to evaluate the reference quality.
• Palmyra Creative, a new language model of the writerIt promises the ability to develop “style” so that the entire output generated by AI does not sound in decline the same.
• During training, AI collects human data biases. When humans interact with AI, there is a Feedback loop That amplifies those biases.

Programming

• Unicon It may never become one of the 20 (or main 100) programming languages, but it is a descendant of Iconwhich was always my favorite language for chain processing.
• What do captchas mean? When the bots equipped with LLM can successfully complete the tasks established for humans?
• EGUIalong with EframeIt is a GUI library and a frame for oxide. It is portable and is executed natively (in macOS, Windows, Linux and Android), on the web (using wasm) and in many game engines.
• For the archivist in us: the Manx The project is not an island in the Irish sea or on cats. It is a catalog of manuals for old computers.
• Cerebrec It is a graphic python Framework for deep learning. It is aimed at Python programmers who do not have enough experience to create applications with Pytorch or other AI libraries.
• Github has announced Free access to github co -driver for all current and new users. Free access provides 2,000 code complexes and 50 chat messages per month. They have also added the ability to use the Claude 3.5 sonnet in addition to GPT-4O.
• DevinThe assisted coding tool that claims to admit the development of software from beginning to end, including design and purification, has arrived General availability.
• Json5, also known as “Json for humans“, It is a variant of Json that has been designed for human readability so that it can be written and maintained by hand, for example, in the configuration files.
• AWS has announced Two new significant services: Aurora DSQLwhich is a distributed SQL database, and S3 tableswhich admits Data Lakehouses through Apache Iceberg.
• Autoflow It is an open source tool to create a knowledge chart. It is based on TIDB (a vector database), LlamaINDEX and DSPY.

Security

• Portspof It is a security tool that makes the 65,535 TCP ports look open for valid services. Emulates a valid service in each port. It makes it difficult for an attacker to determine which ports are really open without probe each port.
• We are going to encryptwhich issues the certificates used by websites (and other applications) to prove their identities, has announced short -term certificates that expires after six days. Short duration certificates increase security by minimizing exposure if a private key is compromised.
• Due to the continuous presence of attackers within telecommunications networks, the FBI and CISA of the United States have recommended The use of encrypted communications protocols. (Although they still want rear in encryption systems, which would make them vulnerable to attack).
• TO New Phishing attack Use corrupt words documents to avoid safety verifications. While the documents are corrupt, Word can recover them.
• LLM Rocking It is a new class of attack against language models that prevent railings preventing objectable production from reaching the user. These attacks take advantage of the career conditions in the interaction of the application with the users.
• Bootkitty it’s a Uefi bootkit That goes to Secure Boot in Ubuntu systems. It seems to have been developed by Cybersecurity students in KoreaThen he leaked (possibly accidentally). It has not yet been found in nature, but when it is, it will be a dangerous threat.
• Def with has started a project to Improve cybersecurity for water infrastructure In the United States. They are starting with six water companies that serve rural communities.

Quantum computing

• Google has built to quantum computing chip in which a Logical Qbitic corrected by error It can remain stable for an hour. The “lower threshold” passes: the error rate decreases as physical qubits are added for errors correction. The chip was built in the new Google manufacturing installation.

Web

• Google is adding “store reviews“To Chrome. Reviews are summaries of reports generated by the known sources that report scams and other problems.
• Here is how to do it Creation of transmission text user interfaces On the web. The transmission text is almost a need to build chatbots driven by AI.

Biology

• Yes, we can have a virtual flavor. A research group has developed a Lollipop interface so that people can experience taste in virtual worlds.