Easy AI Daily | January 17, 2026

Products & Applications

• OpenAI launches ChatGPT Go, starts testing ads: OpenAI released ChatGPT Go globally at $8/month — 10x more messages than Free, file uploads, image generation, longer memory/context, and unlimited GPT-5.2 instant usage. Ads will be tested on Free and Go tiers; Plus/Pro/Business/Enterprise remain ad-free. (ChatGPT Go intro, Ads principles)
• Claude Cowork opens to Pro users: Still a research preview, with session renaming and connector improvements. Users report organizing 400+ files consumed 97% of session quota, suggesting limits remain tight for complex tasks. (Reddit)
• Gemini 3 Pro complaints mount: Pro users report degraded performance in long projects — off-topic answers and code corruption — with suspicions Google quietly reduced the context window. Some are switching to GPT-5.2 Thinking or Claude. (Reddit)
• Perplexity Pro quota frustration: The 100/day advanced-query cap is being exhausted within hours by heavy users, prompting cancellation considerations. (Discord)
• Agentic IDE/CLI costs: Reported bills include Cursor Ultra burning 20% quota per orchestrator run, Qoder near $400/month, and Gemini CLI at ~$120/day (10M tokens). Users want clearer usage stats and configurable model tiers per agent. (Cursor Discord)
• LMArena updates: PDF-chat testing underway; image leaderboards refreshed, with FLUX.2 klein ranking well in mid-to-high positions. (Image Edit, Text-to-Image)
• Hawk Ultra hype: Movement Labs' model reportedly generates 9k–20k lines of code from a single prompt; community awaits comparisons with Gemini 3 Pro and Claude Opus. (X)

Models & Capabilities

• Sam Altman teases "Very fast Codex" and memory improvements; developers discuss shifting toward fast-model, human-shepherded workflows. (Tweet)
• Codex CLI supports open-weight models: Ollama enables codex --oss, recommending 32K+ context; Codex also experiments with mid-thinking steering. (Ollama)
• SWE-rebench (Dec 2025): Claude Opus 4.5 leads at 63.3% on 48 new GitHub PR tasks; GPT-5.2 xhigh at 61.5%; Gemini 3 Flash Preview outperforms its Pro sibling; GLM-4.7 is the strongest open-source model. (swe-rebench.com)
• Unsloth long-context RL: Claims 20K context RL training on 24GB VRAM and up to 380K on a 192GB B200 — 7–12x context extension without accuracy loss. (Reddit)
• GLM-Image (Zhipu + Huawei): Multimodal model trained entirely on Ascend 910, supporting 1024–2048 resolution without extra training; strong Chinese text rendering; claims ~60% better tokens-per-joule than H200; API ~¥0.1/image. (Reddit)
• VoxCPM (OpenBMB): Open-source token-free TTS generating continuous waveforms, LoRA-finetunable, ~0.15 real-time factor on a single 4090. (Tweet)
• Translate Gemma launched: Full collection on Hugging Face; Ollama integration available. (HF collection)
• OpenBMB AIR framework: Splits preference data into Annotation/Instruction/Response; reports +5.3 average improvement across 6 benchmarks using 14K curated samples. (Tweet)

Agents & Tooling

• Human-in-the-loop validated again: Same models fail more in fully autonomous runs; a human gatekeeper materially improves reliability. (Discussion)
• Jerry Liu: chunking is dead?: For document sets in the hundreds, agents directly opening files and using ls/grep beat fixed chunk + vector DB RAG; databases only needed at larger scale. (Tweet)
• Parallel tool calling: Claude supports multiple tools per API call with parallel tool-use control; OpenRouter community expects lower latency and cost. (Docs)
• Agent orchestration UI/CLI wave: SpecStory CLI, sled UI, OpenWork (local computer agent with Ollama), and more. (SpecStory)
• Claude Flow v3 skepticism: Claims 75–80% token reduction (2.5x effective Claude Max capacity) via TypeScript+WASM multi-agent swarms; community criticizes lack of benchmarks. (GitHub)

Infrastructure & Hardware

• "Inference explosion year": A widely-shared post argues agents raise I/O ratios making prefill the dominant cost; context caching becomes essential; prefill/decoding disaggregation worsens utilization and requires redesigned scheduling/memory hierarchies.
• SambaNova SN40L runs DeepSeek R1: Higher throughput than NVIDIA configurations at high concurrency; ~269 tokens/s single-user peak. (Benchmark)
• Epoch AI: ~30GW global AI datacenter capacity, comparable to New York State's summer peak — an upper bound, not actual load. (Tweet)
• Kernel engineering: NVIDIA's CuTe/cuTile approaching cuBLAS performance with cleaner block-level code; AMD gfx942 multi-L2 coherence requires manual buffer_inv sc1 invalidation. (ROCm docs)
• PCIe/power management matter: A 3090 in Gen3 x1 dropped inference from 120 to 90 t/s; sleep(2s) in benchmarks causes GPU downclocking. (GPU MODE)
• GPU market: A $500 "broken" A100 40GB turned out fine (add active cooling); RTX 5070 Ti discontinued and 5060 Ti 16GB production cut, raising prices on a budget-70B staple. (Reddit)

Research & Methods

• Mamba-2 rewrote its core algorithm (parallel scan → block-diagonal GEMM) to lift Tensor Core utilization from 10–20% to 60–70%; meanwhile Microsoft's RetNet was displaced by its own Phi Transformers — illustrating the "Transformer attractor" of hardware + architecture + resources. (Essay)
• Multi-vector retrieval resurgence: A 32M-parameter model with ColBERT/ColPali-style multi-vector retrieval can approach 8B-model quality. (Tweet)
• "Information Gravity" hallucination theory with hysteresis firewalls — more thought experiment than validated method. (GitHub)
• MMLU-Pro fixes: EleutherAI patched the dataset and lm-evaluation-harness; old MMLU-Pro scores may be skewed and should be rerun. (PR)

Industry & Policy

• OpenAI monetization: Ads on Free/Go plus tiered subscriptions mark a shift toward a hybrid ads+subscription model for 900M weekly users.
• Higgsfield AI: $130M Series A at $1.3B valuation, claiming $200M ARR within 9 months. (X)
• Tax automation startup: $3.5M seed (General Catalyst et al.) to make US personal tax filing free and one-click. (Founder tweet)
• Anthropic Economic Index #4: Decomposes AI usage into "economic primitives" — task complexity, education level, autonomy, success rate — for finer labor-market analysis. (Report)
• OpenAI ad principles: Answers won't be ad-influenced, ads clearly labeled, conversations not shared with advertisers — though communities worry about long-term incentive drift.
• Jailbreaking research: Active work in BASI Discord on NSFW bypasses, refusal inversion on Llama3, and URL/OCR injection filtering tricks; vendors patching continuously.
• ZKP for AI moderation: A proposal to prove "content passed safety review" via zero-knowledge proofs without revealing content itself.
• Machine consciousness at AAAI 2026: Dedicated workshop by CIMC; submissions due January 23. (Details)

📌 Source: Easy AI Daily