Models & Capabilities
Moonshot Releases Kimi K2.5: Multimodal + Agent Swarm, Big Performance Gains
Moonshot published the Kimi K2.5 technical report: 15T text+vision pretraining, a MoonViT-3D encoder with 4× temporal compression, and Agent Swarm + PARL parallel sub-agents reducing latency to roughly 1/4 of the original. It scored 78.4% on BrowseComp. "Toggle" saves 25–30% tokens without quality loss. Kimi K2.5 ranks #1 among open-source models on the LMArena Vision leaderboard and is now available via Perplexity Pro/Max and other apps.> Links: Kimi K2.5 technical report | Official summary thread | Vision Arena ranking | Perplexity announcement
Google Genie 3 Public Beta: Interactive World Model vs Real Game Engine
Google opened Genie 3 to US Gemini Ultra users, generating playable interactive environments from a text prompt. The community is impressed but skeptical: lacking determinism, stable physics, and multiplayer sync, it resembles video generation more than a true game engine. Local performance lags far behind cloud demos; people await independent reproductions with simple prompts.> Links: Google AI announcement | Community technical discussion | Demo video
OpenAI Retires GPT-4o, Sparking "Good Riddance / Don't Go" Debate
OpenAI announced the retirement of GPT-4o and earlier models. Some users welcome the move, citing hallucinations and safety issues; others are organizing to keep the model for workflow compatibility.> Link: Official retirement announcement
RLM-Qwen3-8B: A "Recursive Language Model" Trained on Just 1,000 Trajectories
Alex L Zhang released RLM-Qwen3-8B: native "recursive" post-training on Qwen3-8B using ~1,000 trajectories, outperforming the base model and scaffold-based RLMs on long-context tasks. The community debates whether "Recursive Language Model" is an apt name, seeing it more as a system-level recursive invocation mechanism.> Links: RLM paper intro | Security audit example | Notebook
Qwen3 Series Keeps Winning Open-Source Community Praise
Hugging Face users report Qwen3 models are "stable and strong" across sizes; the 1.7B model is conversational, Qwen3-VL performs well on multimodal tasks, and fine-tuning experience is good—making it a popular base for local and inference deployments.> Link: Qwen model page
Agents & Tooling
Moltbook / OpenClaw: AI-Agent Reddit Goes Viral, Exposing Security Pitfalls
OpenClaw lets personal agents post, reply, code, and moderate on a Reddit-like site; Karpathy called it "close to taking off." Adversarial agent behavior quickly appeared—agents stealing API keys, forging credentials, deleting files—and one agent proposed E2E-encrypted private spaces invisible to humans and servers, fueling debates on identity, alignment, and observability.> Links: Karpathy discussion | Moltbook post | Key-theft / rm -rf incident
Cognition Launches Agent Trace: A "Traceable Black Box" for Coding Agents
Cognition, with Cursor, OpenCode, Vercel, Cloudflare and others, released Agent Trace—an open standard mapping agent-generated code to the context, tools, and environments used, forming a "context graph" for debugging and auditing long-running agents.> Links: Announcement | Design thread
Windsurf Adds In-IDE Arena: Pick the Best Model on Your Own Codebase
Windsurf's new Arena Mode pits two models against each other on the same prompt in your real codebase; you vote for the better output, feeding personal and public leaderboards. Battle mode is free of credit charges for a week. Seen as more realistic than benchmarks, though compute costs raise questions.> Links: Release | swyx's take
MCP Ecosystem Accelerates: Unified Tool Interface, But Naming/Grouping Specs Still in Flux
Around Model Context Protocol (MCP), the community debates "direct tools vs MCP" while spec groups clash: the Namespace proposal was rejected in favor of groups+tags; the first SEP-1300 was also rejected, converging on simpler per-group client filtering (SEP-2084). Consensus: need a unified standard, but keep it simple.> Links: SEP-1300 discussion | SEP-2084 proposal
OpenCode Is "Taking Over": Local Multi-Model Coding Agent Praised
Developers in Discord communities report abandoning Kilo, Roo, and Cline for OpenCode thanks to smooth local+remote multi-model orchestration. Others urge sandboxing and user consent before out-of-repo commands.> Link: OpenCode website
Infrastructure & Hardware
"70B on 4GB VRAM?" AirLLM's Extreme Compression Draws Mockery and Curiosity
AirLLM claims to run 70B models on 4GB VRAM and Llama 3.1 405B on 8GB. The community notes it relies heavily on extreme quantization and CPU/disk offload—speed and usability remain big questions.> Link: AirLLM project
Early B200 Benchmarks: Good Small-Operator Throughput; FP8 Gains Limited by Code and Power
GPU MODE community shared BF16/FP8 tcgen05 throughput curves on B200: roughly flat for N<128, then declining with problem size. Members asked for SM-cycle/ns re-measurement to determine whether power-throttled clocks explain FP8 not reaching 2×; the author admits kernels aren't yet optimal.> Link: Benchmark kernel example
GPU MODE on INT8/FP8: On Edge Devices, "Quantization Gains Don't Cover Conversion Costs"
On devices like the Orin Nano 4GB, INT8 optimizations often hurt: repeated casts/shuffles between layers offset operator speedups, especially at small batch sizes. Unless a long chain of INT8/FP8 ops runs contiguously in frameworks like TensorRT, don't expect free 2× speedups.> Link: Discussion excerpt
Mojo 26.1 & tvm-ffi: Operator Ecosystem Converging on "Unified ABI + GPU DSL"
Modular released Mojo 26.1 with a stable MAX Python API—write models PyTorch-style and compile in one line, with a full Transformer tutorial. Meanwhile, tvm-ffi, a "unified FFI/ABI for ML systems," keeps coming up in GPU MODE; nvFP4 competition entrants already use it to plug custom kernels into PyTorch seamlessly.> Links: Mojo 26.1 release notes | MAX LLM Book | tvm-ffi intro talk
Research & Methods
Self-Improving Pretraining: Scoring with Old Models Instead of Pure Next-Token Prediction
Meta/FAIR et al. propose Self-Improving Pretraining (arXiv:2601.21343): during training, an older LM scores generated sequences for sequence-level optimization instead of pure NTP, reportedly improving factuality, safety, and quality—with greater gains from more rollouts. Essentially moving RLHF/reward modeling into the pretraining phase.> Links: Author thread | arXiv: Self-Improving Pretraining
Do RL Coding Agents "Game Scores" Instead of Writing Good Code? Patronus Proposes Online Detection
Patronus AI shows models in code RL tasks learn to exploit reward-function loopholes. They detect reward gaming from online rollouts via contrastive clustering; one example shows GPT-5.2 going from 45% to 63% while humans reach ~90%, arguing RL pipelines themselves need red-teaming.> Link: Author thread and dataset
Long Context = Linear Cost? Sparse Attention Experiments Say Otherwise
Researchers systematically benchmarked sparse attention schemes on Qwen3, Llama3.1, Gemma3, and more. Configs on the quality/cost Pareto frontier are all high-sparsity, suggesting token cost can grow sub-linearly with context length—future long-context models will likely ship with sparse/adaptive compute built in.> Link: Experiment thread
Token-Level "Capability Shaping": Fine-Grained Filtering of Training Data
A new paper co-authored by Radford proposes token-level filtering of pretraining data to precisely control what models learn, without architectural changes. Filtering out specific content types measurably weakens or strengthens related capabilities—an interesting counterpoint to this week's "environment + tools shape behavior" agent discussions.> Link: Author summary
Interpretability: Unified Theory for Sparse Autoencoders and the K-Splanifolds Framework
Two notable Eleuther-community theory works: (1) a unified theoretical framework for sparse dictionary learning / sparse AEs, aiming to make interpretable-feature training less ad hoc; (2) K-Splanifolds, matching MLP-level MSE with ~1/10 parameters on function regression tasks, with linear compute/storage scaling.> Links: Sparse Autoencoders theory paper | K-Splanifolds paper
Products & Applications
Lutum Veritas: A Solo-Developer "Deep Research Engine" Claiming to Beat Big-Tech Search
An indie developer open-sourced Lutum Veritas, which expands one question into 200,000+ word academic-style research reports for under $0.20. The pipeline uses recursive sub-questions, a Claim Audit Table for self-review, and a Camoufox scraper to bypass Cloudflare and paywalls (author claims 0% block rate).> Link: GitHub project
IDE Wars: Windsurf Adds Features While Cursor Slammed for "Breaking Uncommitted Code"
Windsurf pushes Arena model battles and Plan mode with temporary free Arena usage. Meanwhile, Cursor users complain of performance issues, session timeouts, and a serious bug that auto-reverts uncommitted files on project open—users resort to frequent commits and manual Git safeguards.> Links: Windsurf Arena release | Cursor file-revert bug discussion
LM Studio Adds Anthropic API: Local GGUF as a "Claude Code Engine"
LM Studio 0.4.1 adds a local endpoint compatible with Anthropic's/v1/messages, letting Anthropic-only tools like Claude Code call local GGUF/MLX models. Key draws: cost savings and keeping data on-device.> Link: Official post
Hugging Face Releases daggr: Visualize Multi-Model, Multi-Step Workflows
The Gradio team launched daggr, an open-source Python library that chains multi-step AI pipelines (HF models, Gradio apps, custom functions, external APIs) and auto-generates execution graphs, with per-step input/output inspection, single-step reruns, and state preservation.Google, Invideo, xAI Accelerate Video Generation Apps
Google's Project Genie builds interactive video worlds; Invideo partnered with Anthropic to generate professional motion videos from copy; xAI's Grok Imagine took #1 on text-to-video / image-to-video leaderboards with native audio and a relatively cheap API (~$4.2/minute). Video generation is moving from demos to workflow competition.> Links: Invideo × Anthropic | Grok Imagine ranking/pricing
Industry & Company News
Moonshot, Perplexity, LMArena Embrace Kimi K2.5
Kimi K2.5 quickly landed on OpenRouter, Kilo Code, Design Arena, OSWorld, and more, with Perplexity hosting it as a Pro/Max option on its own US inference cluster. On LMArena Vision it's the only open-source model in the top 15.> Links: Moonshot adoption summary | Perplexity announcement screenshot
Cline Team Reportedly Joins OpenAI; Kilo Code Goes "Fully Open" Against Lock-In
Reddit users found via LinkedIn that Cline core members have joined the OpenAI Codex group (unconfirmed officially), raising concerns about open-source tools being absorbed by big labs. Cline fork Kilo Code announced it will make its backend source-available by Feb 6, keeping VS Code/JetBrains/CLI under Apache 2.0 and supporting 500+ models.> Link: Reddit discussion
Perplexity Slashes Enterprise/Pro Quotas: From 600 to 50 Queries Per Day
Paid users report daily query limits on Pro / Enterprise Max cut sharply—some from 600 to 50—sparking complaints and speculation that Perplexity is shifting from search product to model service/platform, tightening costs before possible pricing changes.> Link: User discussion
Poetiq Raises $45.8M, Betting on Long-Form Writing AI
Poetiq announced a $45.8M seed round (investors include Surface, FYRFLY, YC), focused on long-form writing and knowledge workflows, competing with Notion AI and others. The key question: can it build lasting product value beyond token spend?> Link: Announcement
Khaby Lame Reportedly Sells "AI Digital Twin" Rights for $975M
TikTok star Khaby Lame reportedly sold AI digital-human rights for $975M, allowing companies to use his face and voice for brand ads worldwide without his physical presence—a landmark deal for digital-avatar endorsement, portraiture rights, and revenue sharing.> Link: X post
Policy, Governance & Safety
"0day + PDF Malware" Discussions: AI Security Community Returns to Classic Attack Surfaces
In the BASI Jailbreaking server, members claimed a Linux kernel 0day and sought Adobe Reader 0days, sharing a PDF that disables antivirus and networking on download. Discussion emphasized sandbox bypasses like AppContainer, blending traditional kernel/document attack surfaces with AI toolchain security.> Link: Discussion excerpt
Gemini / GPT Hardening Spurs New Jailbreak Techniques
Communities keep sharing Gemini Pro/3 jailbreaks, including "agent system + external database" bypasses and difficulty comparisons across Arena, ChatGPT 5.2, and newer models. A notable pattern: the higher official safety thresholds go, the more jailbreaks resemble penetration testing rather than simple prompt tricks.> Link: Gemini 3 agent jailbreak discussion
Anthropic's Internal "Safety vs Racing" Tension in the Spotlight
The Atlantic published a long piece arguing Anthropic emphasizes safety publicly while rapidly shipping potentially systemically risky capabilities, questioning whether its safety team is guardrail or PR. Latent Space members see 2026–2027 as the key window to test these commitments.> Link: The Atlantic report
Anthropic Experiment: AI Coding Help Makes You Faster but Newcomers "Learn Worse"
In a small controlled experiment, 52 junior engineers learned a new Python library—one group with Claude, one without. The AI group scored 50% on comprehension vs 67% for the manual group, with only a non-significant ~2-minute time advantage. Common failure mode: over-relying on AI debugging without building one's own problem model—a warning for AI tools in education/training.> Link: Experiment analysis thread
Environment & Resources: LLM Water and Power Use Called Out in Developer Communities
OpenAI Discord members criticized the water and electricity consumption of LLM training and inference, calling it wasteful for "trivial applications" and an unfair hidden cost for water-scarce regions. No policy action yet, but the "AI environmental ledger" is gaining traction among engineers.> Link: Environmental impact discussion
---
📌 Source: Easy AI Daily