Easy AI Daily | January 17, 2026

Key points

Products & Applications

• OpenAI launches ChatGPT Go ($8/month) — 10x more messages than Free, file uploads, image generation, longer memory/context, and unlimited GPT-5.2 instant. Ads will be tested on Free and Go tiers; Plus/Pro/Business/Enterprise remain ad-free. (Announcement | Ad principles)
• Claude Cowork opens to Pro users — Still a research preview; users report organizing 400+ files consumed 97% of session quota, sparking complaints about tight limits. (Reddit discussion)
• Gemini 3 Pro performance complaints — Pro users report degraded long-project performance and suspect a silently reduced context window; some are migrating to GPT-5.2 Thinking or Claude. (Reddit feedback)
• Perplexity Pro's 100/day advanced-query cap — Heavy users exhaust quotas within hours and are considering switching to token-billed alternatives.
• AI IDE/CLI costs under fire — Reported bills: Cursor Ultra burning 20% of quota per orchestrator run, Qoder near $400/month, Gemini CLI ~$120/day at 10M tokens. Users want clearer usage dashboards and hybrid model routing.
• LMArena adds PDF chat testing and updated image leaderboards; Hawk Ultra (Movement Labs) is hyped for generating 9k–20k lines of code per prompt as a possible "Opus killer."

Models & Capabilities

• Sam Altman teases "Very fast Codex" and improved ChatGPT memory; developers discuss workflows shifting to fast models with human-in-the-loop steering. (Tweet)
• Codex CLI supports open-weight models via codex --oss (Ollama), with 32K+ context recommended and mid-generation steering being tested.
• SWE-rebench (Dec 2025): Claude Opus 4.5 leads at 63.3%, GPT-5.2 xhigh at 61.5%; Gemini 3 Flash Preview beats its own Pro; GLM-4.7 is the strongest open-source model. (swe-rebench.com)
• Unsloth claims 7–12x longer RL context (20K on 24GB VRAM, up to 380K on a 192GB B200) via data migration and new batching algorithms.
• Zhipu & Huawei release GLM-Image — trained fully on Ascend 910, supports 1024–2048 resolutions without extra training, strong Chinese text rendering, ~60% better tokens-per-joule than H200 (claimed), API ~0.1 RMB/image.
• VoxCPM (OpenBMB) — open-source token-free streaming TTS with LoRA fine-tuning, ~0.15 real-time factor on a single 4090.
• Translate Gemma now on Hugging Face and integrated into Ollama; OpenBMB AIR alignment framework reports +5.3 average points across 6 benchmarks using 14K curated samples.

Agents & Tooling

• Human-in-the-loop validated again as a major reliability multiplier for agentic workflows.
• Jerry Liu (LlamaIndex): fixed chunking + vector DB RAG is dying for small document sets — direct file tools (ls/grep) beat pre-chunked embedding until scale demands a database.
• Claude and OpenRouter now support multiple parallel tool calls in a single request, cutting latency and cost.
• New orchestration tools emerging: SpecStory CLI, sled UI, OpenWork local computer agents; Claude Flow v3 claims 2.5x effective Claude Max capacity via WASM multi-agent swarms, though the community questions its benchmarks.

Infrastructure & Hardware

• "Inference explosion year": a widely shared essay argues prefill now dominates cost, context caching becomes standard, and scheduling/memory hierarchies need rework.
• SambaNova SN40L runs DeepSeek R1, beating NVIDIA clusters on high-concurrency throughput (~269 tok/s single-user peak).
• Epoch AI estimates ~30GW of installed AI datacenter power — roughly New York State's summer peak.
• Kernel engineering: NVIDIA CuTe/cuTile tiling nearing cuBLAS performance; AMD gfx942 multi-L2 coherence requires buffer_inv sc1 to avoid stale caches.
• PCIe matters: a 3090 on Gen3 x1 drops inference from 120 to 90 tok/s; sleep(2s) in benchmarks causes GPU downclocking artifacts.
• GPU market: a working used A100 40GB found for $500; RTX 5060 Ti 16GB reportedly discontinued/reduced, raising prices.

Research & Methods

• Mamba-2 rewrote its core scan as block-diagonal GEMM, lifting Tensor Core utilization from 10–20% to 60–70%; RetNet's abandonment shows the Transformer–hardware–resources lock-in.
• Multi-vector retrieval (ColBERT/ColPali-style): a 32M-parameter model with multi-vector can approach 8B-model retrieval quality.
• Information Gravity proposal (GPU MODE) models hallucination loops via excitation thresholds with a hysteresis firewall — currently more thought experiment than method.
• MMLU-Pro fixed: EleutherAI patched the dataset and lm-evaluation-harness; older MMLU-Pro scores may be skewed and should be re-run. (PR)

Industry & Business

• OpenAI monetizes ~900M weekly users via ads + tiered subscriptions, seen as a pivot toward a hybrid ad/subscription model.
• Higgsfield AI raised $130M at a $1.3B valuation, claiming $200M annualized revenue in under 9 months.
• A tax-automation startup (Saket Kumar) raised a $3.5M seed from General Catalyst to make US personal tax filing free and one-click.
• Anthropic's 4th Economic Index introduces "economic primitives" (task complexity, education level, autonomy, success rate) to quantify AI's labor-market impact.

Policy, Governance & Safety

• OpenAI ad principles: answers won't be altered by advertisers, ads clearly labeled, conversations not shared with advertisers — but the community worries about long-term incentive drift.
• BASI Jailbreaking community catalogs daily jailbreak techniques (Gemini NSFW bypasses, Llama3 refusal reversal, OCR/cool-link filter evasion) that vendors rapidly patch.
• A zero-knowledge-proof AI moderation proposal would let platforms verify content was screened without exposing the content itself.
• AAAI 2026 will host a Machine Consciousness workshop (CIMC), with submissions due January 23, focusing on operational detection methods rather than philosophy.

📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant