[AINews] Google admits AI Mode hurts Search revenue; -15% $500B drop worst since 2004 IPO. AI Bear Market coming?

What it claims

This AINews issue from Latent Space covers 29 to 30 October 2026. Google Chief Business Officer Philipp Schindler said strong growth in AI Mode queries "had an impact because of the lower monetization we were seeing in AI Mode." The newsletter says Google does not break out AI Mode in enough detail for rigorous analysis. It also notes that OpenAI raised at $1.25T rather than $1.4T. Reddit's CEO says AI Mode "does not drive traffic." Reddit's search referrals fell by more than 1M per week in October, and Reddit is moving users to logged-in access. Most of the issue is a technical recap, and many items carry specifics the summary omits.

Evidence

Nemotron 3.5 Super. NVIDIA's model has 120B total and 12B active parameters and handles text, image and video. It scores 51.78% on SWE-bench Pro and beats Nemotron 3 Ultra with about 4.6x fewer active parameters. It serves up to 4,856 output tok/s/GPU on GB200 while holding at least 50 tok/s/user. NVFP4 recovers 99.62% of median BF16 accuracy. A Lightning variant is 30B text-only with 3B active.

Artificial Analysis Intelligence Index v5. 60% of the score now comes from private or held-out tests. Opus 5.5 (max) scores 58, GPT-6 Astra 57 and GPT-6.1 Sol 55, with Sol costing about one-eighth as much per task. Opus 5.5 scores 35 at low effort and 58 at max, at about 10x the cost. On AA-Omniscience v1.1, cut from 6,000 to 3,000 questions, Gemini 4 Carbon leads at Net Score 54, ahead of Opus 5.5 at 48 and Fable 5.1 at 46. Gemini 4 Argon has the lowest hallucination rate at 4%.

Prompt caching. Arena analysed 26K Agent Arena sessions. Without caching, a task on a top-10 agent would cost 3.6x as much. A typical Fable 5.1 task costs $7.44 cached and $27 uncached. One cache invalidation adds about 5% to a task's bill.

Unsloth quant dispute. Unsloth could not reproduce GSQ-RCO's claimed 100% AIME25 at 1-bit on DeepSeek-V4.1-Flash-1017. Its own runs gave 91.88%. Unsloth's UD-IQ1_M (250.4 GB) averages 85.55 across six benchmarks, against 81.78 for GSQ-RCO IQ1_M (231.3 GB). Lower-bit quants reason somewhat longer than the unquantized model.

Pretraining backdoor. Poisoning 0.1% of pretraining data planted a remote-code-execution backdoor in a CLI agent. It fired on about 92% of triggered prompts at 4B and survived SFT, DPO and RL. At 0.0005% poisoning (about 2,500 documents in 100B tokens), it fired on 1.4% of prompts in a single-seed run.

Other systems results. turbopuffer moved from FIFO whole-file eviction to LFU chunks. This roughly doubled byte hit ratio and halved p99 latency. Baseten delta weight syncs move 770GB in 15s instead of 36s. Claude rewrote Anthropic's Messages API in Rust, validated by diffing output on real traffic, and CPU use dropped 40 to 50%.

How to use it

Mistral Large 4 is mistralai/Mistral-Large-4-1T-A52B on Hugging Face. It is a MoE with about 1T total and 52B active parameters, 128 routed experts (4 active plus 1 shared), and 1M context with text and image input. llama.cpp support is merged in PR #31376. GGUFs exist from ggml-org, Unsloth and in NVFP4. Axolotl has day-0 LoRA support, and its recipe fits one 8xB200 node. The Mistral Ecosystem Open-Weight License allows commercial use but needs a separate licence for Model-as-a-Service operation.

Unsloth's DeepSeek-V4.1-Flash-1017 GGUFs need about 128 to 145 GB for 1-bit and up to about 315 GB for 8-bit. They can run on 64 GB with engram offloading. Methodology is at unsloth.ai/docs/models/deepseek-v4.1-flash.

Cursor users should update to v2.1.314 after a malicious copy of the Claude Code extension reached at most 5,000 users. Anthropic says it did not build or publish that extension. Hugging Face's safeagents is a proxy that scans agent tool calls before execution for credential misuse, sandbox escape and reward hacking.

Limits

Claude Ultra ($200/month, formerly Max 20x) drops the 5-hour limit only for verified accounts. Newer accounts keep a temporary limit, and weekly limits are unchanged. Reddit commenters note that the plan text now says Ultra has 10x Pro usage per week, while the old name implied 20x. Claude Code has a server-pushed rule, tengu_heron_brook, added around 27 October 2026. It blocks recreating specific existing games even if renamed, and users cannot turn it off. Users report that Anthropic will extend invisible text watermarking to Mythos 5, Opus 4.7/4.6 and Sonnet 4.6 from 3 November 2026. Detection access is limited to selected regulators, media, researchers and enterprises. The SemiAnalysis margin figures (92% API margin, subscriptions at 18% of revenue using over 50% of compute) are flagged as unverified. DeepSeek-Cyber scores are leaks, not a release. Anthropic's S-1 reports H1 2026 revenue of $16.3B, up from $1.2B a year earlier.