Easy AI Daily | January 10, 2026
A full English translation of the zhichai.net AI daily digest covering models, agents, infrastructure, research, products, industry news, and policy.
Models & Capabilities
DeepSeek releases MHC training method; V4 rumored to focus on coding
DeepSeek proposed Manifold-Constrained Hyper-Connections (MHC) in a paper, using constrained matrices to solve signal amplification and gradient explosion issues when training a 27B ultra-high-connectivity model — seen as a small but practical improvement for stabilizing deep, highly connected architectures. Meanwhile, multiple sources report DeepSeek V4 is expected in February, focused on long-code understanding and generation, with internal benchmarks reportedly beating GPT/Claude on coding, though it has not yet launched.> Links: MHC technical analysis | V4 coding report (The Information) | V4 February launch (Reuters)
Falcon-H1R-7B: a small reasoning model joins the open race
TII (UAE) released Falcon-H1R-7B, an open 7B-class reasoning-oriented model that performed strongly on Humanity's Last Exam, τ²-Bench Telecom, and IFBench. Drawback: its license requires attribution, lowering its "openness" score. It is viewed as a strong contender in the small-model reasoning track, pressuring closed large models.> Links: Performance & openness analysis | Open-model competition discussion
FineTranslations: 1 trillion tokens of parallel corpus via Gemma3
Researchers used Gemma3-27B to translate FineWeb2 multilingual data into English, building FineTranslations, a 1T+ token parallel corpus suitable for multilingual alignment, distillation, translation/RAG training, and evaluation.> Link: Dataset introduction
LTX-2: open-source audio-video generation under 8GB VRAM
LTX-2 is an open-source audio+video generation model that can produce videos up to ~20 seconds on under 8GB VRAM; ~5 minutes for 20 seconds on a 4090-class GPU, with LoRA training code provided. Practical for local short-video and A/V LoRA work.> Links: Model page | Reddit discussion
LFM 2.5 small model challenges big models on synthetic data quality
Community testing shows LFM 2.5B approaches Qwen3 30B Q8 and Qwen3 235B Q3_XL when generating training synthetic data. LFM2.5-1.2B-Instruct-SDG with a lightweight frontend runs local SDG pipelines in ~1GB VRAM.> Link: LFM2.5-1.2B-Instruct-SDG
Agents & Tooling
MCP ecosystem accelerates: official servers, mcp-cli, new implementers
The OpenAI camp shipped an official MCP server unifying docs, guides, and AppsSDK into a tool surface consumable by IDEs/agents; the community built mcp-cli, replacing verbose tool descriptions with discovery and claiming ~99% token savings. New developers are now implementing the spec and filing GitHub issues — the protocol has moved from "reading docs" to "real implementations."> Links: OpenAI MCP server | mcp-cli | Spec implementation issue
Anthropic skill.md + industry-wide "skills" systems form a standard agent abstraction
Anthropic's blog defines skills via skill.md: a Markdown description plus scripts/data files that agents read on demand, avoiding stuffing all tool descriptions into prompts. GitHub Copilot, Claude Code, and Cline all promote skills — versionable, on-demand instruction packages — becoming a common abstraction layer for agent development.> Links: Anthropic agent skills | VS Code Agent Skills | Cline skills compatibility
AI21 uses git worktree as "transactional workspaces" for multi-agent concurrency
AI21's MCP Workspace provides init/clone/compare/merge/delete primitives; each concurrent attempt gets an independent git worktree directory, running up to 16 candidate solutions simultaneously before merging the winner — a real application of transactions + concurrency control to code workspaces.> Link: AI21 workspace thread
InfiAgent and file-based workspaces: long-horizon agents hinge on "context engineering"
InfiAgent stores long-term state in files/directories, rebuilding only a snapshot plus a fixed recent-interaction window each step to avoid unbounded context growth. Combined with community discussion of "agent drift," context/file management is increasingly seen as the real challenge for long-horizon agents, not the models themselves.> Links: InfiAgent summary | Agent drift discussion
OpenRouter routing upgrade: performance floors and skills loading
OpenRouter added "partition" sorting for auto-routing with performance floors, selecting sufficiently fast models without added latency; a Provider Explorer shows model coverage per cloud. The SDK gained a Skills Loader to wrap Anthropic-style skills as context usable by any model.> Links: Routing docs | Provider Explorer | Skills Loader
Open-source RAG Demo Toolkit: brand-agnostic chat assistant via OpenRouter
A developer open-sourced an Agentic RAG Demo Toolkit built on OpenRouter API + FastAPI + Qdrant: drop in your docs and logo to get a complete RAG bot with retrieval, frontend, and demo video.> Links: Toolkit repo | Demo video
Claude Code skills system and code-simplifier agent open-sourced — but maturity questioned
Anthropic open-sourced its internal "code simplification agent" for handling huge PRs to the official plugins repo, and 2.1 added recursive skill fork orchestration with per-sub-agent context. Early users report issues: sometimes deleting important code or odd refactors, showing auto-refactor agents still need extensive evaluation and guardrails.> Links: code-simplifier plugin | Reddit discussion
Infrastructure & Hardware
AI compute doubles every 7 months; single datacenters near 1GW
Epoch AI estimates global AI compute doubles roughly every 7 months based on accelerator shipments, with NVIDIA taking 60%+ of new compute. Anthropic's Indiana datacenter is estimated at ~750MW, soon reaching 1GW-class. Compute and power constraints now directly shape product strategy and freebie-abuse controls.> Links: Compute growth analysis | Anthropic datacenter estimate
Modal shares ops experience with 20,000 concurrent GPUs
Modal detailed managing 20k+ GPUs across clouds and 1M+ total instances, covering public-cloud GPU failure rates, dropouts, and scheduling pitfalls — health checks, multi-cloud redundancy, and scheduling strategy are must-haves for large-scale inference/training platforms.> Link: Modal engineering thread
PyTorch 2.9 fixes flash_attn_varlen + compile incompatibility; ~50% measured speedup
GPU MODE community reports frequent graph breaks with flash_attn_varlen + torch.compile on torch 2.4; upgrading to 2.9 eliminated the problem with ~50% throughput gains. Nightly builds add better variable-length sequence APIs.> Link: Community discussion
ROCm and AMD GPUs: raw compute fine, ecosystem remains the barrier
Users report AMD 7900XTX under Vulkan/ROCm: compute close to a 4090, but CUDA ecosystem remains ~10% faster overall, plus issues like PyTorch misidentifying iGPUs and ROCm 7.1 autotuning problems. Verdict: viable for budget/VRAM-focused builds if you accept toolchain friction.Consumer GPU choices: T4/A2000, 7900XTX, RTX 5090 trade-offs
Community advice: under 75W power limits consider Tesla T4 / RTX A2000; the 7900XTX is attractive on price/VRAM but depends on Vulkan/ROCm; the RTX 5090 reportedly draws a 400W VBIOS minimum — unsuited for a quiet big-VRAM workstation. Compute is no longer scarce; power and ecosystem are the main constraints.Research & Methods
GDPO: multi-reward RL algorithm patching GRPO's "advantage collapse"
GDPO (Group reward-Decoupled Normalization Policy Optimization) normalizes each reward separately, avoiding GRPO compressing different reward combinations into identical advantage values, yielding more stable multi-objective RL convergence — possibly explaining many observed GRPO instabilities.> Links: GDPO thread | GRPO critique
LM Jigsaw: multimodal models "fall off a cliff" at 5×5 puzzles
A new benchmark has VLMs reconstruct images by swapping puzzle tiles: frontier models solve ~95% at 3×3 but nearly 0% at 5×5, with token cost rising from ~55k to ~345k. Authors argue current models do patch matching rather than true global spatial reasoning — a warning for robotics and navigation.> Links: Project page | GitHub | Demo
CGGR training method: skip gradients on some steps to save 75% memory
An Eleuther community proposal skips gradients on partial steps; training SmolLM-135M on fineweb-edu reportedly uses 6–7GB at batch=4, comparable to normal batch=1, with up to 75% theoretical VRAM savings. Preliminary experiments only."Dead salmon" paper strikes again: many interpretability methods explain random networks too
A preprint, *Dead Salmon: An Artifact of Random Initialization* (arXiv:2512.18792), shows feature attribution, probes, sparse autoencoders, and even causal analysis produce plausible-looking explanations for randomly initialized networks.DeepSeek MHC discussion: scalability needs structural constraints, not blind width
Community discussion: piling on depth/cross-layer connections causes signal explosion/convergence failure; geometric or convex constraints are needed to "leash" information flow. Analogized to ResNet-era tricks — not revolutionary, but likely to become default practice.> Link: Reddit ML discussion
Products & Applications
OpenAI launches healthcare ChatGPT offering; major hospitals onboard
"OpenAI for Healthcare" emphasizes HIPAA compliance and medical customization, live at AdventHealth, UCSF, and Memorial Sloan Kettering. Goals: standardizing clinical documentation, patient communication, and summarization; physician AI usage reportedly nearly doubled in a year.> Link: Official announcement
Claude Code as a "full-stack partner": weekend POCs previously feared
Engineers on Reddit report shipping complex projects over a weekend using Claude Code's multi-turn refactoring, bottleneck prediction, and code review. Caveats: AI accelerates implementation but exposes flawed architectures — humans still own complexity reduction.> Link: Success story discussion
LTX-2, VeridisQuo, Synthia: generation, detection, and synthetic data covered
LTX-2 lowers open-source A/V generation barriers; VeridisQuo on HuggingFace detects deepfakes with GradCAM heatmaps; Synthia generates text with LFM2.5-1B on 1GB VRAM. The generation–detection–synthetic-data chain is complete.> Links: VeridisQuo | LTX-2
Lovable: one system-prompt revision saves $20M in cloud costs
Lovable engineers report that system prompt optimization (less verbosity, clearer task boundaries) improved inference speed ~4% with better design quality — worth roughly $20M/year in LLM costs at their scale. Prompt engineering at product scale is purely a cost problem.> Link: Cost optimization post
Local AI practice: multi-GPU desktop platforms to low-VRAM synthetic data tools
Eloquent, a 9-month hobby project, is a local multimodal platform with multi-GPU sharding, story tracking, and persona evaluation. Community advice: high concurrency needs GPUs; old servers suit 7B-or-smaller deeply quantized models; RAG must tolerate imperfect memory.> Link: Eloquent
Industry & Business
Anthropic tightens Claude Max usage in third-party apps
Anthropic restricted users from using Claude subscriptions in external clients, cutting off some competing products. Developers now treat vendor consumer plans as revocable perks and default to multi-model wrappers and BYO-key architectures.> Links: Developer feedback | Discussion
MiniMax IPOs in Hong Kong, positioned as multimodal "open ecosystem"
MiniMax listed on the HKEX, emphasizing its early bet on unified multimodal (text/voice/video) models and developer programs. IPO cash strengthens its position in China's model arms race.> Link: IPO coverage
China's $22T household savings seen as potential AI "ammunition"
Analysis: Chinese households hold ~$22T in savings, historically only 5% in financial markets; another 5% would release $1T. Combined with Qwen/DeepSeek's open-source price-performance, some expect a "80% capability at 10% price" strategy undercutting Western closed subscriptions.> Link: Discussion
Protege AI raises $30M to supply real-world data for LLMs
Founded 2024, Protege AI raised $30M led by a16z to provide cross-industry multimodal "real-world data" — essentially productizing and API-izing datasets for model developers who don't want in-house data teams.> Link: Funding announcement
OpenAI for-profit conversion lawsuit cleared for jury trial
A US judge ruled a lawsuit over OpenAI's shift from nonprofit to capped-profit structure can proceed to jury trial, potentially exposing internal governance records and setting legal precedent for how AI companies are structured.> Link: Case report
LM Arena data: top-ranked models hold #1 for only ~35 days on average
LMArena statistics show the #1 model stays top for about 35 days on average and typically falls out of the top five within 5 months — auto-routing, fast evaluation, and migration capability matter more than clinging to one model.> Link: Statistics tweet
Policy, Governance & Safety
Radware discloses first "zero-click" server-side ChatGPT vulnerability
Radware reports a zero-click server-side flaw: a specially crafted email could cause ChatGPT's backend to leak sensitive data with no victim interaction. Disclosed to OpenAI. Enterprises should watch such "ordinary input" attack surfaces.> Link: Radware advisory
OpenAI/Anthropic billing and quota disputes erode trust
Users report OpenAI auto-upgrading $20 Plus plans to $200 Pro with difficult refunds; Claude Code Pro users saw unexplained usage spikes, suspecting background requests. Community suggests virtual-card limits; opaque usage accounting is a trust risk.> Links: OpenAI billing thread | Claude Code usage thread
NO FAKES Act voice fingerprinting may lock down TTS creation
Critics warn the proposed US NO FAKES Act's fingerprinting provisions could require non-removable watermarks on all speech synthesis, legally blocking public voice cloning; supporters stress protecting artists and individuals from unauthorized synthesis.> Link: Reddit discussion
Deepfake risk rises: chatbot-linked deaths entry and detection tools
A Wikipedia entry, Deaths linked to chatbots, collects extreme cases; tools like VeridisQuo visualize suspicious regions. Core debate: responsibility among vendors, users, and platforms, and whether regulators should codify generation liability.Anthropic publishes practical agent evals guide: start from real failure traces
Anthropic's blog distinguishes capability vs. regression evals, introduces code/model/human graders and pass@k vs pass^k metrics, and recommends building eval sets from real production failure traces — iterating instructions, tools, and evals together to avoid "strong on paper, failing in production."> Link: Demystifying evals for AI agents
---
📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant