Easy AI Daily Digest – January 17, 2026: ChatGPT Go Ads, GLM-Image, Inference Year

Forum topic · 小凯 · 2026-03-27

Summary

This January 17, 2026 edition of the Easy AI Daily covers major AI industry developments. OpenAI launched ChatGPT Go at $8/month and began testing ads on free and Go tiers, pledging that responses won't be influenced by advertisers. Anthropic opened Claude Cowork to Pro users, though users report heavy usage limits. Developers complained about Gemini 3 Pro performance degradation, Perplexity Pro's 100/day limit, and costly agentic IDEs like Cursor Ultra and Gemini CLI. On models, SWE-rebench ranks Claude Opus 4.5 first (63.3%), with GLM-4.7 the strongest open-source entry. Zhipu and Huawei released GLM-Image trained on Ascend 910 chips, while OpenBMB open-sourced VoxCPM token-free TTS and Unsloth demonstrated 7–12x longer RL context training. Infrastructure stories include SambaNova SN40L outperforming NVIDIA on DeepSeek R1 throughput, Epoch AI estimating ~30GW global AI datacenter capacity, and Mamba-2's Tensor Core rewrite. Business news features Higgsfield's $130M raise at a $1.3B valuation and Anthropic's fourth Economic Index report on AI's labor market impact.

Product & Application Launches

OpenAI launches ChatGPT Go, begins testing ads on free and low tiers

OpenAI rolled out ChatGPT Go globally at $8/month—offering 10x more messages than free, file uploads, image generation, longer memory/context, and unlimited GPT-5.2 instant access. It also announced ads will be tested on Free and Go tiers; Plus/Pro/Business/Enterprise remain ad-free.

Links: ChatGPT Go intro | Ad principles | Go plan announcement

Claude Cowork opens to Pro users, usage pressure is real

Anthropic opened Claude Cowork (still research preview) to Pro users with session renaming and connector improvements. One user reported that batch-organizing 400+ files consumed 97% of session quota, suggesting limits remain tight for complex tasks.

Link: Reddit discussion

Gemini 3 Pro 'unusable' complaints, suspected context window shrink

Pro users report Gemini 3 Pro degradation on long projects—off-topic answers, garbled code—with some suspecting Google quietly reduced context length. Heavy users are moving to GPT-5.2 Thinking and Claude.

Link: Reddit feedback

Perplexity Pro's 100/day advanced queries deemed too few

Power users exhaust the 100 daily advanced-model requests within hours, prompting cancellations and switches to token-billed alternatives.

Agentic IDEs/CLIs burning money

Users shared bills: Cursor Ultra's single orchestrator run ate 20% of quota; Qoder hit ~$400/month; Gemini CLI burned 10M tokens (~$120) in a day. Community wants clearer usage stats, caps, and small-model-subagent/large-model-main-agent configuration.

LMArena adds PDF chat, image leaderboard updates

LMArena is testing direct PDF chat; FLUX.2 klein series ranks mid-to-high on image editing/text-to-image boards, showing small models catching up quickly. Links: Image Edit | Text-to-Image

Hawk Ultra hyped as an 'Opus killer'

Movement Labs' Hawk Ultra generates 9k–20k lines of code from a single prompt, ideal for scaffolding projects; open-source plans unclear. Link: Movement Labs on X

Models & Capabilities

'Very fast Codex' teased; memory upgrades

Sam Altman hinted at improved ChatGPT memory and an incoming "Very fast Codex," sparking discussion of fast-model + human-shepherding workflows over slow one-shot answers.

Codex CLI supports open-weight models and longer context

Ollama enabled open models via codex --oss, recommending 32K+ context. Codex is also experimenting with mid-thought steering without interrupting generation. Links: Ollama OSS support | Context length

SWE-rebench: Claude 4.5 leads, GLM-4.7 strongest open model

On 48 new GitHub PR tasks: Claude Opus 4.5 tops at 63.3%, GPT-5.2 xhigh at 61.5%. Gemini 3 Flash Preview beats its own Pro despite smaller size. GLM-4.7 is the strongest open model, near GPT-5.1-codex level. Links: Reddit summary | SWE-rebench

Unsloth pushes RL context to hundreds of thousands of tokens

Via data migration and new batching, Unsloth claims RL training to 20K context on 24GB VRAM and up to 380K on a 192GB B200—a 7–12x extension without precision loss. Link: Reddit discussion

Zhipu & Huawei release GLM-Image, trained on Ascend 910

Multimodal model trained entirely on Ascend 910, supporting 1024–2048 resolution without extra training, emphasizing Chinese text rendering; claims ~60% better tokens-per-joule than H200. API priced ~0.1 yuan per image. Link: Reddit discussion

VoxCPM: open-source token-free real-time voice cloning TTS

OpenBMB's VoxCPM generates continuous waveforms rather than discrete audio tokens, supports LoRA fine-tuning, and streams at ~0.15 real-time factor on one 4090. Link: Announcement

Translate Gemma is live

Google's Translate Gemma collection is on Hugging Face with Ollama integration and recommended prompt formats. Links: HF collection | Ollama integration

OpenBMB AIR alignment framework

Splits preference data into Annotation/Instruction/Response; reports +5.3 average points across 6 benchmarks using 14K curated samples. Link: Announcement

Agents & Tooling

Human-in-the-loop reconfirmed as a reliability multiplier

Engineers report that fully automated runs fail often, while a human gating uncertain outputs dramatically improves results—the gap between the two curves is human-in-the-loop value.

Jerry Liu: file tools over traditional RAG

LlamaIndex's Jerry Liu argues fixed chunking + vector DBs are what's dead; at hundreds-of-docs scale, agents using ls/grep and on-demand context expansion are simpler and more robust. Databases are only needed at larger scale. Link: Thread

Claude and OpenRouter support parallel multi-tool calls in one request

Parallel tool use reduces round trips, latency, and cost—an important agent orchestration capability. Link: Docs

Agent orchestration UIs/CLIs emerging

SpecStory CLI, sled UI, OpenWork (local computer agent with Ollama) and others are launching rapidly. Links: SpecStory | sled | OpenWork

Claude Flow v3 claims 2.5x Claude Max value; skepticism abounds

Rewritten in TypeScript+WASM with multi-agent swarms and shared memory, claiming 75–80% token reduction. Critics cite marketing buzzwords and missing benchmarks. Link: GitHub

Infrastructure & Hardware

'Training has peaked; the real pain is inference'

A widely shared essay argues agents raise I/O ratios making prefill the dominant cost, context caching becomes standard, and prefill/decoding separation worsens device utilization—requiring redesigned scheduling and memory hierarchies.

SambaNova SN40L runs DeepSeek R1, beats NVIDIA on throughput

At high concurrency SN40L showed higher throughput with ~269 tok/s peak per user; per-hour pricing wasn't published, but non-NVIDIA inference is now competitive. Link: Benchmark

Epoch AI: global AI datacenter installed capacity near 30GW

Rough estimate via GPU sales × rated power × 2.5 facility factor—comparable to New York State's summer peak demand (an upper bound, not actual load). Link: Estimate

CUDA/ROCm kernel engineering

NVIDIA's CuTe/cuTile enables near-cuBLAS performance with block-level code and improved swizzling; AMD gfx942 multi-L2 coherence requires manual buffer_inv sc1 to avoid ghost cache errors across XCDs. Link: ROCm gfx942 docs

PCIe and power management matter more than expected

A 3090 on Gen3 x1 dropped inference from 120 to 90 t/s; a sleep(2s) in benchmarks causes GPU downclocking, inflating measured latency. Check board wiring and power policy before benchmarking.

Consumer/used GPU market

A $500 "broken" A100 40GB turned out fine (add active cooling); RTX 5070 Ti discontinued and 5060 Ti 16GB production slashed, raising prices for the budget 16GB option.

Research & Methods

Mamba-2 rewrote its core for Tensor Cores; RetNet abandoned

Mamba-2 replaced parallel scan with block-diagonal GEMM, lifting Tensor Core utilization from 10–20% to 60–70%—a concession to NVIDIA hardware. Microsoft's RetNet was quickly displaced by dense Phi Transformers, showing the architecture+hardware+resources lock-in of the Transformer. Link: The Transformer Attractor

Multi-vector retrieval resurgence

Experiments show a 32M-parameter model with ColBERT/ColPali-style multi-vector retrieval can approach 8B-model quality—putting complexity into index structure rather than model scale. Link: Experiments

'Information Gravity' + hysteresis firewall for hallucinations

A speculative GPU MODE proposal treats token selection as excitation flow with a hysteresis firewall resetting state at threshold—more thought experiment than method. Link: GitHub

MMLU-Pro dataset and harness fixed

EleutherAI patched TIGER-Lab/MMLU-Pro issues in lm-evaluation-harness; old MMLU-Pro scores may be biased—rerun before comparing. Links: PR | Dataset discussion

Industry & Business

OpenAI monetizes 900M weekly users: ads + finer subscription tiers mark a turn toward a hybrid ads/subscription model. Commentary
Higgsfield AI: $130M Series A at $1.3B valuation, claiming $200M ARR within 9 months. Announcement
Tax automation startup: $3.5M seed led by General Catalyst 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

Policy, Governance & Safety

OpenAI ad principles: no answer changes, no conversation leaks—but long-term incentive drift concerns

OpenAI pledges ads won't affect answers, will be clearly labeled, and conversations won't be shared with advertisers. The community worries about gradual drift in recommendation ordering and default tools as revenue pressure grows.

BASI community actively probing jailbreaks

Summaries include Gemini NSFW jailbreaks ("free but short-lived"), Grok called among the wildest models, refusal inversion on Llama 3, and URL/text-filter bypasses via cold links and OCR injection; vendors are patching continuously.

ZKP for privacy-friendly AI content moderation

A proposal: a unified classifier flags violations, then a zero-knowledge proof attests content passed moderation without revealing it—letting platforms enforce moderation without snooping.

Machine consciousness on AAAI 2026 agenda

AAAI 2026 will host a CIMC workshop on machine consciousness: defining/detecting AI consciousness, behavior vs. internal states, and ethical implications. Submission deadline January 23; organizers want concrete methods, not philosophy. Link: Workshop note

---

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