From Chatbot to Work Assistant: Six Months of AI Advances, Gaps, and Real User Experiences

1. AI Model Release Cadence and US‑China Competition

Stanford 2026 AI Index reports 12 core model version releases between Dec 2025 and Jan 2026, average interval 4.8 days. This rapid iteration turned models into multimodal assistants.

Anthropic, xAI, Google, OpenAI, Alibaba, DeepSeek occupy the same score band; by Mar 2026 the performance gap between leading US and Chinese models narrowed to ~2.7 %.

Alibaba’s Qwen (Tongyi Qianwen) family exceeded 1 billion downloads, becoming the world’s most‑downloaded open‑source model; after summer 2025 China overtook the US in overall model performance.

2. Strengths in Coding vs Weaknesses in Document Generation

OpenAI CEO Sam Altman stated at a 2026 seminar that GPT‑5.2 excels at coding and reasoning but performs poorly on writing tasks because development resources prioritized inference and coding.

Anthropic disclosed in May 2026 that about 90 % of its internal code is generated by AI; employees shifted from typing to supervising AI outputs, reducing a multi‑hour report to an initial draft in roughly half an hour.

For typical workers, AI gains in code generation and data calculation are offset by frequent failures on seemingly simple tasks, limiting replacement of experienced staff.

3. Hallucination and Information Staleness

Multimodal generation quality reached 87 % of expert human level; language understanding accuracy rose to 92 %. However, new visual‑reasoning models showed a 28 % hallucination rate despite a 40 % improvement in tool‑calling ability.

An experiment demonstrated that a large language model could be convinced of a completely false statement for a week at a cost of $12.

“Lies with eyes open” – Google AI incorrectly reported that a Japanese department store would close in Feb 2026, requiring urgent correction.

Information “shelf life” – over 80 % of surveyed enterprises reported encountering AI‑generated misinformation.

Useless work – an ACL 2026 paper from Nanjing University described “blind self‑thinking” where models produce verbose but irrelevant output when prompts are vague.

4. Impact on Enterprise Roles

Gartner 2026 data shows AI‑enabled agents in enterprise applications grew from < 5 % in 2025 to 40 % in 2026 (eight‑fold increase).

Replaceability rates reported: legal document drafting 92 %, junior assistants 87 %, basic market analysts 85 %, news editors 81 %.

Workers experience a mix of fascination and frustration: AI still produces “idiotic” errors while expanding into new domains.

Code example

END
如果看到这里，说明你喜欢这篇文章，请
推
荐、转
发、点赞
。同时
标星（置顶）
本博主可以第一时间接受到博文推送。
私信
说出你想要的资源给你安排!!!
👆
长按上方二维码 2 秒
说出你想要的资源
我知道你 “
在看
”