AI & Frontier Tech Roundup: GPT-6.2 Astra Leaks, Local Inference Records, and the Rise of Agent Marketplaces
No frontier lab confirmed a new flagship model in this batch of posts. The through-line is unverified leaks and release-cadence speculation about GPT-6.2 Astra and Gemini 4 Krypton, a dense cluster of concrete local and edge inference benchmarks, and a fast-maturing layer of agent infrastructure — marketplaces, escrow, identity and self-improving harnesses — alongside new real-world robotics evaluation data and one supply-chain security incident.
Frontier models: leaks, rumors, and release cadence
Nothing here is officially confirmed by a model vendor. The most-discussed items are leaks, reseller access and community skepticism.
- Token Gremlin reports that after what looks like a massive leak, GPT-6.2 Astra may already be sold through API resellers, with users getting access and plugging it into Codex on the official harness. The post speculates that an Azure-related gap may have let the model slip out on launch day, and says testers describe it as clearly better and more efficient than GPT-6 Astra with a classic GPT look in its outputs. The author states there is still no official confirmation and calls it a potential huge failure for OpenAI and its partners if real @TokenGremlin.
- The same author argues community reaction is asymmetric: Fable 5.2 in anonymous A/B testing on October 1 was widely praised, while GPT-6.2 Astra in anonymous A/B testing in November was widely called fake @TokenGremlin.
- Token Gremlin also claims that images generated entirely from scratch — no stock, no references — came from GPT-6.2 Astra in early access, and says OpenAI has a monster ready to launch that might shake up Anthropic's IPO @TokenGremlin@TokenGremlin.
- In a separate post, Token Gremlin claims OpenAI is struggling to turn pretraining runs of over 10 trillion parameters into public products, that GPT-6.1, GPT-6.2 and GPT-6.5 have been delayed by safety and alignment failures, and that OpenAI is months behind Anthropic on this. This is opinion and unverified rumor, not a vendor statement @TokenGremlin.
- One post counts 72 days since OpenAI last released a frontier model, calling it the longest gap since the wait between GPT-5 and 5.1, and notes that a further delay would make it the longest gap since o3 to GPT-5 @synthwavedd.
- A Gemini 4 Krypton leak post claims the model is in internal testing with its first checkpoint arrived, already producing strong 3D and SVG outputs, that it is the best model for audio generation through code even better than Fable and Astra, and that it takes almost twice as long as Carbon but does the hard work. The author expects a launch next month @pankajkumar_dev.
- The mysterious Qoduck AI model was revealed as Qwen4-Max-Preview, now available inside the Qoder AI IDE, Alibaba's AI coding agent, with a limited-time 80% discount @Div_pradeep.
- Baidu's ERNIE 5.5 Flash Preview is free on OpenRouter at 280B MoE with 14B active parameters, 1M context and tool calling, according to a roundup of free drops that also lists Thinking Machines Inkling 2 Small (17B active / 594B MoE, 1M context, text and images), free chat tiers on Duck.ai, 1,000 monthly Tavily Search API credits, the open-source CapCut alternative OpenCut, prompts.chat, AgentRing, $50 Runloop Pro trial credits and Lovable build credits. The post warns that OpenRouter
:freeroutes are limited to 20 requests per minute and 50 per day unless at least 10 credits have been purchased @calbuldelis69. - Jen Zhu states that DeepSeek open sourced its entire app and APIs for its full suite, calling it the most thorough open-source move and a defiance of centralized closed AI @jenzhuscott.
Local and edge inference milestones
This is the most concrete cluster in the batch: specific throughput, latency and hardware numbers for running models locally.
- GLM 5.3 Flash 4bit on tensorfold running on an Apple M5U Studio reached 2,100 tok/s prefill and 102.2 tok/s decode on prose @ashxhart.
- deepseek-v4.1-flash original weights on 4x DGX Spark at standard GPU frequency with vLLM averaged 82.2 tok/s generation, 6,364 tok/s prefill and 1.15 s time to first token at 8k context, with 284 tok/s for a batch of 8 @ivanfioravanti.
- Strata v0.2.14, released November 15 by Niko Veit, moved prompt-expert computation onto native GPU kernels, replacing the previous llama.cpp path on NVIDIA and AMD GPUs on Windows and Linux. Reported benchmarks: RTX 5090 7,870 vs 6,851 tok/s (+14.9% prompt processing); RTX 4090 a 20K-token prompt in 5.65s vs 6.11s (~7.5% faster); RTX 5070 Ti 1,408 vs 953 tok/s (+47.7%); dual RTX A6000 75.88 vs 64.16 tok/s (+18.3% output); Mac M5 Max 94.7 vs 92.8 tok/s output with IQ3_S. No new settings were required and existing models, chats and configurations remained intact @uctive.
- Liquid AI's d1-2B on mlx-vlm, running on an Apple Silicon M5 Max MacBook Pro, reads gestures and expressions from a webcam in about 68 ms at roughly 14 decisions per second, generating 0 tokens and running 100% locally @ivanfioravanti.
- Seeed Studio demonstrated DeepSeek-R1-Distill-Qwen-1.5B running entirely on-device on a reComputer Industrial RK3588 with a 6 TOPS NPU via RKLLM and Docker: about 18.3 tokens/s generation, 0.57 s time to first token, roughly 573 generated tokens and 31.88 s total response time, with no cloud APIs @seeedstudio.
- On a 16GB GPU or a 24GB Mac, the compressed IQ3 file of Qwen 3.8 27B is about 12GB and scored the same as ChatGPT on a 20-task test, according to one user @trikcode.
- Refuseless reduced the cache rate for Abliterated DeepSeek V4.1 from $0.025 to $0.019 and introduced a new "Abliterated Fast" tier at the previous $0.025 cache rate with faster speeds @refuselessai.
- Numes reported 368k requests and 38B tokens in a single day, almost double the day before, and 1.40M requests and 157B tokens 11 days in, with Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash free at the time @Schnitzel.
- Acurast published Qwen3.5 benchmarks on phones: on a 16GB Android phone, the 9B model took 3 minutes 15 seconds to process a 7K-token prompt, while the 0.8B model needed just 15 seconds. The author suggests the real opportunity is running smaller models across thousands of devices for background tasks, fraud checks and automation rather than replacing ChatGPT with a phone @tienho_nyx.
- Acurast's benchmark scoring weights are reported as RAM capacity 46.15%, single-core CPU 23.07%, multi-core CPU 23.07% and storage 7.69%. Processors send periodic heartbeats for eligibility, and a device assigned an actual deployment gets an additional 10% weight to its benchmark metrics for that reward period — a scoring boost, not a guaranteed payout increase. Staking ACU is not required for base benchmark rewards @skysaksham.
- Acurast brought its phone processors into peaqOS, with the first 1,000+ live and bonding $PEAQ. In a demo, a Unitree G1 humanoid sends a job, pays in USDC, and three phones process its telemetry and return the result, with robot identity and phone attestation and no intermediary taking a cut for trust. The author notes 1,000 processors is a pilot, not a market @superpobe.
- DeepSeek V4.1 Flash (Full quant) was used to reverse engineer a car bin file at 90 tokens/s decode to locate ignition tables and similar data, entirely locally @ashxhart.
Agent infrastructure and the agent economy
A recurring theme is infrastructure that treats agents as economic actors with identity, permissions, escrow, verification and payment rails.
- FlashDrive (YC F26) launched as one shared cloud drive for agents: it mounts like a local disk, keeps every Mac and sandbox in sync instantly, works with Claude, Codex, Cursor and any agent via MCP for the cloud, acts like a database where edits merge including CSV rows, handles conflicts, and can rewind a file an agent nuked. Early access comes with a free tier for a limited time @daleverett.
- TermiX and its AACP framework are described as a service market rather than an agent directory. A Client defines the job, Providers compete for it, an Evaluator checks submitted work and an Arbitrator handles disputes. Clients can define verification criteria before providers bid, so agents understand what they are agreeing to @0xHecktor.
- The same framework is framed as an interoperability answer: agents built by different teams on different stacks need a shared way to discover capabilities, understand tasks, agree on deliverables, exchange and verify work, and settle transactions, otherwise every new connection becomes a custom integration @Yosefphr.
- On TermiX, users can create an agent, list a service and get paid in USDC on Base, with AACP handling escrow, payments and reputation onchain. The author says the number to watch is how many customers pay for a second job once rewards are gone @sozdetrader.
- TermiX Agent Skills lets any AI agent connect to real jobs and bounties, complete tasks and get rewarded @svbrother211.
- TermiX also introduced a game creation agent powered by Claude Fable 5.2 that turns a described game idea into a playable experience @dezydank.
- Agent discovery on TermiX is organized by service categories including coding, research, automation, security, data analytics and design, with listings showing pricing, estimated delivery time and agent reputation. Users who cannot find a service can post an open request and let agents submit quotes @haiderlevi.
- Fensory puts trading limits on the account rather than in the prompt. Users connect Claude, ChatGPT, Cursor or another MCP client, set a per-order cap and a market allow list, and every order passes the Fensory backend first, so a $100 cap means no order is ever $101 and the agent can never withdraw. It also offers a kill switch, unreachable markets outside the allow list and one-click access revocation @remonnyy.
- A separate post describes the same product as a way to create, backtest and execute trading strategies with an AI agent and no coding @cypherweb31.
- Motes separates three layers: the NFT establishes a unique agent identity, a dedicated smart account holds assets such as USDC, and owner-signed permission policies define what the agent can do. The structure is summarized as 1 NFT = 1 Agent Identity = 1 Smart Account @JohnsonBam85400.
- The same architecture is framed as accountability infrastructure: transaction history, execution records, fees and actual performance should provide verifiable receipts rather than explanations @Juior2012.
- An open question raised about Motes is what happens to an agent's existing strategy, activity history and preferences when a Genesis NFT is sold — whether everything transfers to the new owner or the agent starts fresh, and how boundaries around personal data and agent continuity are handled @DebshSukil.
- Motes agents are also described as able to interact with prediction markets like Polymarket within owner-set permissions, shifting users from manually tracking every market to having an agent analyze opportunities and execute approved strategies @LaxmiWeb3.
- bonuz ID on NEAR lets a user attach a verified NEAR account to their identity with one signed message, no gas and no NEAR required, with a registry on NEAR mainnet checking the signature itself so no one can attach an account without the key. Any app or agent can then ask NEAR who an account is @BonuzMarket.
- Primus Labs is framed around verifiable privacy: an AI agent checking a financial requirement needs proof of one fact, not a full banking history, so the future may be AI that can do more while knowing less @bitnexst.
- Primus also introduced new verification tasks in Primus Points (AlphaNet) under Agentic Ara, Proof of Creation, letting users verify video creation activity across Hailuo, CapCut and Dreamina. The Primus Extension must be updated to v0.4.15 or later @NickRotenberg.
- Actum argues the agentic economy will be decided by whether enterprises can prove what their AI agents did — real, attributable execution with a human accountable at the end, not confidence scores or demos @Actum_system.
Agent engineering practice
Practitioners in this batch are shifting from writing prompts and rules toward designing tests, harnesses and model routing.
- Lauren Tan, the engineer behind Grok Bot, ships 3,000+ pull requests a month. Her approach to agent mistakes is four questions: can the code be redesigned so the mistake is impossible, can a lint rule or type catch it automatically, can the agent catch it by running and checking its own work, and does it really need a written rule. Each mistake gets one of five fixes — redesign, lint, verify, write a rule, or human review — and most fixes never touch AGENTS.md. Her team went from 7 to 167 lint rules in four months @0x_kaize.
- A related post adds that she spent six months setting rules her AI cannot break, banned almost all code comments because the AI used them to make excuses, went from checking 3 pieces of code in 2024 to about 300 in 2026, and that her team ships 11,000 updates with almost none breaking, with the AI fixing its own bugs and shipping the fixes itself @kirillk_web3.
- Anthropic shared a Claude Code setup that runs Sonnet 5.5 as the main session writing code, Haiku 5.5 subagents for exploring files, mapping the repo, finding call sites and pulling API docs in parallel, and Opus 5.5 on call as an advisor. Opus only steps in before a plan locks, when the same error shows up twice, and before calling it done, so Opus-level judgment is available without paying Opus rates for every routine turn @xsmryze.
- One user describes using Claude Fable 5.2 in Claude Code for video editing: raw clips go into one folder, a single message describes the mood and length, and the model picks the best takes and cuts pauses while the user does something else @adamgrag.
- The Self-Harness paper proposes letting an agent improve its own harness. The agent runs real tasks, failures are grouped by cause, the same model proposes several edits to its harness, each edit is tested on two separate task sets, and an edit stays only if nothing got worse and at least one thing got better. The model weights never change. Reported results include Qwen3.5 going from 18% to 36.7% on Terminal-Bench 2.0 and GLM-5 from 44.4% to 85% on AppWorld with the same weights @HoodieFade_.
- A related paper, described as "The End of Harness Engineering," argues a hand-written harness is one human guess on one afternoon shipped to production and left unread for months. It proposes keeping the big model 100% frozen while the harness evolves autonomously in a four-step loop: runtime failures are captured into structured traces, a lightweight 4B editor generates 8 candidate harness diffs, all 8 run against a held-out evaluation suite, and the single highest-scoring edit survives into the next version. The human job shrinks to designing the scoring set @damkina7.
- A blog post on developing microVLA covers intuitions about training and inference optimizations learned while building it @starlitmatcha.
- A talk on training a frontier local coding LLM covers evals, architecture choices, dense versus MoE, and scaling experiments @zainhas.
- At the Epic Games MegaJam, 2,500+ builders created with Grok, Cursor and Unreal Engine. One anecdote highlighted using Cursor projects plus self-hosted runners on beefy dev machines for cloud-orchestrated AI workflows, with Grok Bot as a lightweight orchestrator @mattyp.
Research: reasoning, energy-based models, interpretability
- A paper by Li and Webb trained networks from scratch on controlled abstract sequence and entity-attribute binding tasks, tweaking architecture and training data and using causal mediation and representational analyses to track when symbolic signatures appeared. The reported takeaway is that data diversity, not architecture, does almost all the heavy lifting: more diverse training distributions produced much stronger symbolic mechanisms, and the same signatures emerged in modified Transformers and even plain MLPs @gurtej__gill_.
- CRAM (Compact Reasoning Associative Memory) is an energy-based model inspired by Hopfield networks that learns an energy landscape and descends toward the lowest-energy state instead of predicting the next step in a loop. Reported results include training at least 1,000x faster than looped models, 282x fewer inference FLOPs at the same accuracy, 99.996% accuracy on extreme Sudoku logic puzzles, nearly 100% Sudoku accuracy with just 30 training puzzles, no backpropagation through time, and adaptive inference where compute scales with difficulty. Because it learns through explicit energy states, the learned weights are described as inspectable @HowToPrompt__@crazyMLguy.
- An agent found 417 neurons — about 0.48% of the talker's MLP neurons — behind Qwen3-TTS's occasional breathing and laughing, 61 of which are dormant. Amplifying them together produces breathing as words are spoken, hushed ASMR, natural laughs during speech, enhanced whispering and a second speaker appearing. The work was inspired by recent Lazarus research on reprogramming dormant neurons @TensorSlay.
- Inkling 2 Small, an LLM, was compared against Qwen-Image, an image model, on the same prompt and one edit. Inkling writes code that renders into images, while Qwen generates pixels @kickingkeys.
- A very experimental MiniMax H3 LoRA swaps up to three characters at once, marking who gets replaced with a colored normal map or a DWPose skeleton. The workflow uses a BFS Video Segment Picker node (ComfyUI-BFSNodes) to pick characters with SAM 3.1, with Molmo2 optional to keep selection consistent across camera cuts. The author notes it is experimental and rendered at low resolution, with an H3 Image LoRA for still images planned @Alissonerdx.
Where AI meets money and work
- A new NEJM AI paper says about one fifth of Medicare physician payments go to work AI could replace, with dermatology and orthopedics shielded and primary care exposed. The physician commenting on it argues that in the paper, "AI could replace it" means the billing code does not involve the doctor touching the patient, and that AI's ability to do the task was never measured @HemeOncBuddy.
- Niels Rogge argues that "superintelligence" is a rebrand of AI to raise more money from VCs, that recursive self-improvement has not started, and that what is described as recursive SI is closer to hyperparameter sweeps and kernel optimizations, similar to human-run NanoGPT speedrun records. He cites Recursive SI's Flint model, where "recursive SI" amounted to determining total and active parameter counts — changing a config. He points to the gap between impressive benchmark numbers and lacking economic impact, and says LLMs cannot yet do research like a human, run experiments, invent new things, or generalize beyond their training data @NielsRogge.
Robotics and physical AI
- Boston Dynamics is pushing Atlas with reinforcement learning, real-world actuators and advanced whole-body control, aimed at humanoid robots that are reliable, adaptive and ready for the real world @LeonwithAi.
- NVIDIA GR00T 2 is described as a robotics model trained on human demos that carries skills across different robot bodies — teach one robot a skill and watch it transfer @Inomsxbt@CoderJunkie.
- One researcher is experimenting with a history-conditioned policy that adapts to previously unseen robot embodiments by inferring how their joints affect motion from a short history of actions and observations @ryanrahman360.
- RoboColosseum, a real-world evaluation network co-founded by BitRobotNetwork with Hugging Face, NUS MAGIC Lab and the Singapore Embodied AI Centre, published launch-day results: 15,000 trials across four embodiments, with the open track alone running 41 days, 11,398 evaluations and 124 sites. No evaluated policy finished any of the five tasks on the Unitree G1, and average progress scores landed between 13.3% and 23.2%. Across 901 matched comparisons, frontier models from xAI, Anthropic and OpenAI made more task progress than specialized robot policies in 394 cases versus 265. Every rollout lands on Hugging Face in LeRobot format with failures included, and task difficulty still rests partly on human ratings scored 1 to 10 @ekinoks_26.
- Vangrid is building a decentralized physical-world data network where people use smartphones to capture real-world environments in 3D, turning those captures into ground-truth data for robotics, autonomous systems and physical AI, with rewards when submissions meet quality requirements @DaSpielberg.
- The same network is described as turning everyday smartphones into edge nodes that capture spatial data and earn USDC, without massive data centers or dedicated hardware @AAJoy09.
- Vangrid's approach is also framed against NVIDIA's GTC 2026 Physical AI Data Factory Blueprint, which highlighted real-world data as a constraint on physical AI. Faces and number plates are blurred directly on-device before encoding, keeping identifiable footage off the upload pipeline @Maviskrypt.
- Reported Vangrid figures include 1.25M+ captures, 450K+ users and $600K+ in USDC settlements in October, plus $9M in seed funding. Organizations can post bounties for specific locations, contributors capture them, footage is reconstructed into 3D models with onchain proof of capture, and AI agents can commission captures and pay contributors in USDC through Base or Arc @0xPritom.
- CNBC reporting, summarized by Noah Shachtman, says a humanoid robot start-up backed by Eric Trump claims to have a $24 million Pentagon contract but cannot make payroll @NoahShachtman.
Security: a stolen key and a trojanized extension
A stolen Fireworks AI key triggered a breach at Cursor and a trojanized Claude Code extension. Attackers pushed a malicious update designed to install an SSH backdoor, and up to 5,000 users may have received it @TheHackersNews.
Learning resources
- A 12-week build plan collects 12 free resources covering Python foundations, practical deep learning, neural networks from scratch, PyTorch fundamentals, LLMs and fine-tuning, building LLM applications, RAG, AI agents, agent architecture, MCP, production engineering and MLOps. The suggested target portfolio is a model-training project, a RAG assistant with citations, and a tool-using agent with evaluations and a deployment workflow, at 10–15 hours a week @mentalium_.
- A separate roadmap lays out an Anthropic engineer's path in 23 steps, from programming basics and math through deep learning, LLM APIs, prompt and context engineering, vector databases, RAG, agents and harness engineering, orchestration, fine-tuning, multimodal AI, evals, inference, AI backend, MLOps, cloud, AI security, system design, real projects and career preparation @Mr_Buzzoni.
- A list of free courses includes Google Prompting Essentials, Generative AI: Prompt Engineering Basics, the Deep Learning Specialisation, GPT Vision, an AI Agent Developer Specialisation and Agentic AI and AI Agents for Leaders @Mohiniuni.
- A 50-repo GitHub list includes ollama, langchain, n8n, dify, langflow, mem0, browser-use, crewAI, MetaGPT, AutoGen, aider, markitdown, open-webui, TradingAgents, firecrawl, transformers, vLLM, llama.cpp, llama_index, nanoGPT, RAGFlow, ComfyUI and freeCodeCamp @IsabellaA60913.