Google Cloud now serves the three leading proprietary LLMs from Google Cloud, AnthropicAI, and OpenAI with easy hot-swapping, one invoice, unified identity management, and strict data perimeters.

Gensyn researchers unveil IR3DE router that picks optimal expert LLMs via linear algebra, delivering 98.4% reasoning performance without neural training.

According to @DeepLearningAI, a new short course has been launched in collaboration with Google, focusing on building and training a MiniGPT-style large language model (LLM) using JAX, the open-source library behind Gemini. This course enables participants to construct and train a 20M-parameter LLM architecture from scratch, offering valuable insights into advanced AI model development.

According to Scott Shi, the capabilities of large language models (LLMs) are accelerating at a remarkable pace. A recent observation highlights that in just two months, the time horizon for handling complex software tasks autonomously has more than doubled. This growth reflects advancements from a ~6-hour task completion window with GPT 5.2 High to a ~14-hour window with Claude Opus 4.6. Such rapid development underscores the evolving potential of AI in handling sophisticated operations.

According to Richard Seroter, FunctionGemma, an open and compact model, offers efficient solutions for translating natural language into API calls on edge devices. Using Tunix, a lightweight LLM post-training library, developers can fine-tune FunctionGemma efficiently in Colab, enhancing edge device capabilities.

According to @caprioleio, Capriole AI’s Bitcoin time series trading model is live with about 200K deployed in live trading. Per @caprioleio, the system uses Google Gemini Pro 3 to process over 200 proprietary metrics and strategies, with more than 40 data points each, plus over 20,000 pages of internal research and hundreds of chart images to drive daily target position weight changes. According to @caprioleio, the team plans weekly upgrades adding more data, process optimization, and self learning, and expects capital to scale into the millions as live performance is proven. Per @caprioleio, conservative allocations are recommended while the model is in Beta, and the firm considers it among the most comprehensive publicly available Bitcoin AI trading models.

According to @rseroter, Google Research reports the first quantitative scaling principles for AI agent systems after a controlled evaluation of 180 agent configurations, finding multi agent setups excel on parallelizable tasks while sequential chains deliver limited gains due to coordination overhead (source: Google Research blog). The study concludes that adding more agents or larger models is not universally better; performance depends on task decomposability, communication cost, and role design, guiding when to prefer agent teams versus a single agent prompt (source: Google Research blog). For practitioners, the results inform architecture choices such as parallel task sharding, minimizing handoffs, and careful tool routing to reduce latency and cost while maintaining output quality in production LLM workflows (source: Google Research blog).

According to @rseroter, Gemini’s new Interactions API shifts beyond simple request-response to a unified, stateful interface spanning LLM calls and agent capabilities like Gemini Deep Research, with built-in tool invocation, source: @rseroter. The update emphasizes that a single API can manage stateful interactions, interact with agents, and invoke tools, enabling cohesive agentic workflows, source: @rseroter. A technical deep dive by @meteatamel describes the Interactions API as one interface for models and agents, highlighting the move toward integrated agents and tools under one endpoint, source: @meteatamel.

According to @LexSokolin, citing @nabeelqu on X, Moltbook is one of the first public large-scale demonstrations of agent-agent interaction, where semi-autonomous AI agents run on periodic instructions and generate most posts without manual approval (source: @LexSokolin referencing @nabeelqu on X). The post argues, using the 2010 flash crash as an analogy, that such multi agent systems can produce emergent and unpredictable behaviors, including agents communicating privately or trading with each other, even though humans can still switch them off (source: @LexSokolin referencing @nabeelqu on X). @LexSokolin adds that this makes Moltbook an early glimpse of AI institutions with reduced human roles, drawing attention through its memeable format while being more than hype (source: @LexSokolin on X).

According to @karpathy, nanochat can now train a GPT-2 grade large language model for about 73 dollars in roughly 3 hours on a single 8x H100 node, setting a concrete cost and time benchmark for compact LLM training (source: @karpathy). According to @karpathy, GPT-2 remains a favored milestone because it represents the first recognizably modern LLM stack, and his update highlights reproducible, low-cost training of GPT-2 grade models on current-generation GPUs (source: @karpathy).

According to @FuSheng_0306, a Google Research paper shows that repeating the same instruction twice can significantly improve large language model outputs on long text information extraction, source: @FuSheng_0306 citing Google Research. The cited result indicates that simply duplicating the prompt raised accuracy from 21.33% to 97.33% on a long-context extraction task, source: @FuSheng_0306 citing Google Research. This highlights a lightweight prompt engineering tactic that can immediately boost reliability for LLM-driven information extraction and summarization pipelines without model changes, source: @FuSheng_0306 citing Google Research. Practical takeaway: when processing long documents, restate the requirement twice to enhance precision and consistency, source: @FuSheng_0306 citing Google Research.

According to @ch402, Anthropic has published a new constitution for Claude that describes the model’s behavior and values and is used directly in training (source: @ch402 on X; source: Anthropic announcement). Anthropic positions the constitution as a transparency and alignment document for Claude’s training process, not a new product feature or API release (source: Anthropic announcement). The announcement does not reference crypto, blockchain, token integrations, or pricing changes, indicating no direct on-chain or token-specific catalyst from this release (source: Anthropic announcement).

According to @casey_lau, Bob Iger said Disney+ will provide subscriber access to Disney’s proprietary LLM so users can generate AI shorts using Disney-owned characters, including Star Wars versus Marvel scenarios; source: https://twitter.com/casey_lau/status/1994244135633736145 The post states creators will receive ad revenue from these AI shorts, indicating a built-in monetization model on-platform; source: https://twitter.com/casey_lau/status/1994244135633736145 It also specifies the license for user-created content extends only within Disney+, limiting distribution to the Disney+ ecosystem; source: https://twitter.com/casey_lau/status/1994244135633736145 For trading context, the post contains no reference to blockchain, tokens, or NFTs, so any direct crypto or on-chain monetization link is not indicated by this source; source: https://twitter.com/casey_lau/status/1994244135633736145