AI & Frontier Tech Roundup: Gemini 4 Carbon, Agent Swarms, and Physical AI
The through-line across these posts is that frontier AI is separating into three practical layers: flagship models and closed-set decision APIs, agent orchestration that coordinates many workers, and physical-AI systems that learn from simulation or real-world data. The most concrete releases are Gemini 4 Carbon and the Gemini Decisions API, while the strongest research signals are looped transformers and cheaper agent harnesses @GoogleAIStudio@sundarpichai@NVIDIAAI@evilmathkid@dair_ai.
Gemini 4 Carbon and the Decisions API set the model agenda
The clearest model launch in these posts is Gemini 4 Carbon, which Google describes as its strongest model with frontier performance in cybersecurity and long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning @GoogleAIStudio. Google says Carbon is live via the Gemini API and AI Studio for paid API and Google AI Ultra subscribers @GoogleAIStudio. Cursor says Carbon is available in Cursor and that at high effort it is the top model on CursorBench at 60.4% @cursor_ai. Databricks announced a day-0 launch where users can get tokens from Unity Gateway @alighodsi. One post claims Carbon is already live in Google search AI Mode and that Gemini 4 Flash, also called Hydrogen, is imminent @AndrewCurran_.
Separately, Google launched the Gemini Decisions API, powered by Gemini Decision 1 @sundarpichai. Instead of generating text token by token, the API scores closed-set choices across up to 1,000 questions in text, images, audio, video, and PDFs in a single pass @sundarpichai. It supports three decision types: yes/no, multiple choice with up to 255 options, and rubric scoring @sundarpichai. Pricing is $0.10 per 1M input tokens with $0 billed for output tokens, which the post says is about 5x cheaper than Flash-Lite @sundarpichai. The API has a 256k context and 150-250 ms end-to-end latency @osanseviero.
Agent orchestration moves from single assistants to swarms and shared workspaces
Agent tooling in these posts is shifting toward coordinated teams and shared workspaces. Google Antigravity announced /swarm, a slash command for orchestrating specialized agents in parallel across the desktop app and CLI; a single prompt becomes a multi-agent workflow where agents plan, build, and verify together on a live message board, and users can mix models for each role @antigravity. Cursor is getting a multiplayer mode where team members and their agents work in a unified workspace built on top of Projects; one post says Claude and Perplexity are working on similar features @testingcatalog.
Google also released A2UI 1.0, an open protocol and framework for generative UI across web, Android, iOS, wearables, TVs, and other surfaces @GoogleCloudTech. Chrome for Developers launched Modern Web Guidance, which plugs into agents such as Antigravity, Cursor, GitHub Copilot, WebStorm, Codex, Kimi, and Grok and steers them toward current web best practices @ChromiumDev. Google Cloud introduced a Professional Agentic Architect certification for building and deploying agents on Google Cloud @googlecloud.
One practitioner argues teams should draw the work first, including steps, handoffs, decisions, and stops, and then count agents, because many teams build agents first and then find they step on each other @distortgeekin.
Local inference gets faster, smaller, and easier to share
Local and open-model work is becoming more practical. Unsloth released what it calls the highest-accuracy Qwen3.8-27B GGUFs, with Dynamic V3.1 GGUFs leading by 5-10% on average across Terminal Bench, DeepSWE, and other tests, and running on 12GB RAM @UnslothAI. One post claims 1600 tokens per second on a single 5090 running Qwen 3.8 27B @satellitedown. Another argues that on an M5 Max or M5 Ultra, a low-bit DeepSeek 4.1 Flash 1017 is a better choice than Qwen 3.8 27B dense @liuliu.
Omarchy 4.6 will let users share any model running on any Omarchy machine with another machine on their tailnet, so a DGX Spark can serve inference to an XPS13 netbook @dhh. A separate post claims a Qwen 27B model was made smarter without training by rewriting 9 weight matrices, cutting thinking by 27-54%, adding 13 points on LiveCodeBench and 4.3 on GPQA Diamond, and shipping as Apache-2.0 open weights in a 16GB GGUF @nahid_pro09.
For self-hosting, one comparison says vLLM started with PagedAttention and is the broad, safe choice for many models, hardware types, and LoRA adapters, while SGLang started with RadixAttention and reuses shared prefixes for chat, agents, and RAG; the practical rule is to try SGLang for repeated prompts and multi-step flows and start with vLLM for wide model and hardware support @techwith_ram.
Looped transformers and energy-based reasoning lead the research signals
Looped transformers are producing strong results at small parameter counts. Déjà View is a 117M-parameter model that turns photos taken from different angles into 3D geometry and camera positions; by looping a single transformer block, it matches or outperforms models up to 10x larger across five benchmarks with comparable or less compute, and researchers can dial the number of loops to fit their compute budget @NVIDIAAI. A separate entry reports 37% on ARC-AGI-2 Kaggle without LLMs using a 100M looped transformer trained from scratch; the author says it is likely the highest score without an LLM, peaked at 6th place, and beats last year's winning LLM, with looping added in the last week moving the score from 33% to 35% to 37% @evilmathkid.
Energy-Based Reasoners take a different route to cheaper reasoning. They define an explicit energy function over latent states, turning reasoning into gradient descent down that energy surface; the gradient guides refinement, the energy value acts as an intrinsic verifier across candidate solutions, and a flattening gradient gives an automatic halting signal @gurtej__gill_. The reported result is 99.9% accuracy on Sudoku-Extreme with 345x less inference compute than the next-best baseline, plus natural handling of multi-solution problems @gurtej__gill_.
A pretraining note argues pretraining cannot just watch NLL and continued pretraining is not just learning-rate annealing; injecting multi-turn tool patterns in CPT raises the pre-RL starting point @SmartPig_Joe.
Agent cost and evaluation depend on the harness
Agent economics are driven as much by the harness as by the model. UC Berkeley researchers found that harness changes such as prompts and tools barely move accuracy but change cost by up to 2.6x on the same model and task; they identified five ways to cut cost, each reducing context read or number of calls, and built Agency, a layer between harnesses like Claude Code and Codex and the model @dair_ai. Five proof-of-concept techniques cut cost by up to 35% with marginal accuracy loss @dair_ai.
Meta's AutoRefine keeps LLM code judges calibrated by updating judge prompts with signals teams already have, such as whether tests passed and whether a change was merged; on Meta's deployed coding agent, agreement between the judge and real outcomes rose from 0.16 to 0.51 @omarsar0. A separate benchmark, the Voice Arena Dictation Bench, played 761 recordings from 87 speakers into 10 dictation apps on a Mac and used 5 LLM judges to ask whether each draft could be sent as-is; Wispr Flow scored 79% and Apple scored 35% @shobhitbanga.
Evaluation is also being rethought for robotics. One post argues that high benchmark scores do not automatically mean a robot is ready for the real world because scores depend on the task, robot, protocol, and number of trials; the shift in embodied AI evaluation is from leaderboard ranks to understanding real capability @EvalRobo. In peer review, one researcher says LLM reviews for ICLR are inconsistent: they list great results and minor weaknesses, then reject, and LLMs are not fit to judge the puzzle of scientific contributions beyond single instances @HildeKuehne.
Security: tiny backdoors and formal verification
Security posts show both attack and defense. Microsoft researchers showed that changing just 7 neurons out of more than 12,000 in an AI model can plant a hidden backdoor; a hidden trigger made the model generate attacker-chosen commands that an AI agent could execute with its own permissions @TheHackersNews. On the defense side, Google DeepMind proved key security properties of the machine code underlying the Signal messaging protocol, which it calls the first time the actual production code of a widely deployed protocol has been formally verified at this depth; the work used AlphaProof Code, an agentic framework for formal verification running on LLMs including Gemini Carbon, and took a month rather than decades @pushmeet. Google DeepMind is open-sourcing the formal specifications and verification artifacts @pushmeet.
Physical AI: simulation, data, and real-world throughput
Robotics work in these posts centers on data, simulation, and evaluation. Google DeepMind open-sourced MuJoCo Playground, which lets researchers train robots in a physics simulator before deploying them; examples include humanoid robots learning to walk, quadrupeds recovering from falls, and robotic hands manipulating objects, with trained behaviors transferable to physical robots and some simulations runnable in a browser @albacnt. Nebius released Physical AI, an open-source CLI and SDK for teaching robots on Nebius AI Cloud, with examples such as a robot arm picking something up, a humanoid walking, and a quadruped running @nebiusai. PAN2 is pretrained on millions of hours of ego-centric, robotic, and human-motion data and is positioned as a foundation for physical intelligence and robotics @zhiyangx11.
Tesla showed Optimus walking through its office and factory inside a world model at CoRL, with the real policy making every move; if those generated worlds match reality, Tesla can test scenarios without touching a robot @TheHumanoidLabs. Gobano Robotics reports an autonomous towel-folding station with 99.5% task success, 68% of a human worker's throughput on an 18-hour shift, and about 5% throughput improvement per day driven by reinforcement learning; it tracks task success and autonomy separately because a recovered miss differs from a stoppage that pulls a worker over @lukas_m_ziegler. 4D Labs is turning distributed human participation into training data for physical AI, using first-person activity capture, vision-plus-touch data, teleoperation corrections, and world models, with contributors receiving verifiable onchain rights @KunZinCrypto.
Education, work, and the agent-driven lifecycle
A widely shared post describes an MIT committee report that spent five months studying AI's effect on learning and found study groups disappearing, office hours emptying, and problem sets no longer proving anything because AI can produce credible answers; it calls the pattern 'cognitive surrender' @kyronis_talks. The post says 46% of MIT undergraduates use LLMs daily and 90% worry about their own overreliance, and that more students believe AI makes them replaceable than believe it makes them capable @kyronis_talks. Instead of AI detectors or lockdown browsers, the report reportedly recommends oral exams, semester-long portfolios, in-person project work, a required social component in every subject, and possibly rethinking grades @kyronis_talks.
On work, a post quoting Anthropic CEO Dario Amodei says AI will eliminate jobs, probably a large number of them, with entry-level law, finance, and consulting hit first and 50% of entry-level white-collar jobs disrupted in one to five years @shikamarutsai. His advice in that post is to learn to use AI and build a harness that runs it, not just use it directly @shikamarutsai. Another post argues that agent-enabled coding is becoming a new software development lifecycle, with agents writing, refactoring, and testing code, multiple tasks running in parallel, requirements evolving dynamically, and feedback loops happening in real time; it cites agentic coding reports saying teams at Wiz and CRED doubled execution speed @jianw851.
Agent data, privacy, and simulation
New products are turning agent traces and personal context into assets. Brevis launched Arca, a free app that lets users sell AI chats and coding sessions from their computer; users pick sessions and get a price range, and Arca removes identifiable information locally before upload @testingcatalog. Arca can read sessions from Claude Code, Codex, Cursor, and Claude Desktop on Mac and Windows, and is now generally available @testingcatalog.
QVAC released BotPsy, a 0.8B model that sets alarms, moves meetings, and answers messages by calling the right tool; the argument is that an agent which reads email, calendar, messages, and files knows more about a user's life than most friends, so users should choose where it runs, and a model small enough for a phone is a practical answer @qvac. Plurai Simulation is live and free for AI engineers: it recreates a real agent, its tools, and environment from code, adds simulated users and tasks, runs auto evals, attaches verified fixes to failures, and replays production failures against new agent versions inside Claude Code, Codex CLI, or GitHub Copilot @ilan_kadar.
Open media models and creative tooling
Open media models are being used for fast generation and creative workflows. MiniMax says its H3 model, open-sourced 100 days earlier, now has 1,472 models built on it and more than 148M downloads on Hugging Face @RyanLeeMiniMax. pi-H3 generates 10 seconds of 768p video with audio in 9.2 seconds on a single B200 using direct pixel output in 2 or 4 steps and no video VAE decoder @HanshengCh. One post shows H3-Image turning a simple calculator into a full CAD engineering drawing in one shot @0x_zyven.
Creative agent tooling is also expanding. LottieFiles' Creator MCP lets an agent turn an existing SVG logo into a logo-reveal animation, preview it in chat, make a tweak, and export to MP4 inside Claude, ChatGPT, Codex, Cursor, or VS Code @heyfredds. A separate post describes a motion-design workflow with Opus 5.5 and Claude Code that goes from brief to preview to scoped edit to render to reuse @renproton.