Skip to content

Commit 0031b2f

Browse files
yanglbmeyanglbme
committed
docs: add weekly news EN 2026-w21
1 parent 29144e6 commit 0031b2f

1 file changed

Lines changed: 168 additions & 0 deletions

File tree

‎src/content/en/weekly/2026-w21.md‎

Lines changed: 168 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,168 @@
1+
---
2+
title: "Awesome AI Weekly | 2026-W21"
3+
date: "2026-w21"
4+
tags: ["DeepSeek", "Anthropic", "Claude", "Microsoft", "Alibaba", "Google", "NVIDIA", "AI Agent", "AI Pricing"]
5+
summary: "DeepSeek makes 75% discount permanent, Anthropic warns Claude Mythos finds bugs faster than devs can fix them, Microsoft releases Webwright and Fara1.5 agent frameworks, Google AI glasses hands-on"
6+
---
7+
8+
# This Week in Focus: AI Price War Goes Permanent, Agent Frameworks Flood In, Claude Outpaces Developer Bug-Fixing Speed
9+
10+
> Extracting real signal from the noise
11+
12+
---
13+
14+
This week brought a massive wave of developments across multiple dimensions of the AI industry. DeepSeek announced its 75% discount is now permanent, pricing output tokens at least 34x below GPT-5.5 — formally ushering in the AI price war era. Anthropic's Claude Mythos Preview is now finding bugs faster than developers can patch them — both a capability milestone and a security alarm. Microsoft released two major AI agent frameworks, Webwright and Fara1.5, attacking both terminal and browser scenarios simultaneously. And Google's AI glasses hands-on review concluded they're "almost there."
15+
16+
Here are the stories worth your time.
17+
18+
---
19+
20+
## DeepSeek Makes 75% Discount Permanent, Output Tokens Priced 34x Below GPT-5.5
21+
22+
On May 23, The Decoder reported that DeepSeek is making its 75% discount permanent, with output token pricing at least 34x lower than GPT-5.5.
23+
24+
This move signals that DeepSeek no longer treats low pricing as a short-term customer acquisition strategy — it's now a long-term competitive position. While GPT-5.5 and Claude Mythos dominate the performance frontier, DeepSeek is competing through extreme cost efficiency.
25+
26+
> **Awesome AI View:** The AI price war has officially begun. DeepSeek's logic is clear: once model performance gaps narrow to a certain threshold, price becomes the decisive competitive factor. For developers and enterprises, this means AI inference costs will continue to fall as a baseline trend. But for the industry, it may compress margins and force more players to find differentiated paths. Price wars signal industry maturation, but they also accelerate the exit of smaller players.
27+
28+
Source: [The Decoder](https://the-decoder.com/deepseek-makes-its-75-percent-discount-permanent-pricing-output-tokens-at-least-34x-below-gpt-5-5/)
29+
30+
---
31+
32+
## Anthropic Warns Claude Mythos Preview Finds Bugs Faster Than Developers Can Patch Them
33+
34+
On May 23, The Decoder reported that Anthropic issued a warning: Claude Mythos Preview now discovers vulnerabilities faster than developers can fix them.
35+
36+
This isn't marketing copy — it's a genuine engineering bottleneck. When AI's bug-finding capability exceeds human patching capacity, the development process itself needs to be redesigned.
37+
38+
> **Awesome AI View:** This is a fascinating "capability spillover" moment. Security scanners outpacing human remediation isn't new, but when AI can not only find bugs but understand their context and propose fixes, the nature of the problem changes. This could push "AI-assisted remediation" or even "AI auto-patching" into standard practice. It also raises a new security philosophy question: when AI can autonomously discover vulnerabilities, who controls this double-edged sword?
39+
40+
Source: [The Decoder](https://the-decoder.com/anthropic-warns-claude-mythos-preview-finds-bugs-faster-than-developers-can-patch-them/)
41+
42+
---
43+
44+
## Alibaba's Latest AI Model Ran Autonomously for 35 Hours to Optimize Code for Its Own Custom Chip
45+
46+
On May 23, The Decoder reported that Alibaba's latest AI model ran autonomously for 35 hours, optimizing code for its own custom chip.
47+
48+
This marks a rapidly increasing level of AI autonomy in hardware design. AI is no longer just an assistant — it can complete complex engineering tasks independently over extended periods.
49+
50+
> **Awesome AI View:** 35 hours of autonomous operation means AI can transcend human work-hour limitations, iterating and optimizing continuously. This "around-the-clock" engineering capability is key to AI making a substantive impact in high-complexity domains like chip design. Alibaba using AI to optimize its own chip code shows that AI for Engineering has moved from concept to actual productivity.
51+
52+
Source: [The Decoder](https://the-decoder.com/alibabas-latest-ai-model-ran-autonomously-for-35-hours-to-optimize-code-for-its-own-custom-chip/)
53+
54+
---
55+
56+
## Anthropic May Keep Supplying Claude to NSA Despite Pentagon Supply Chain Risk Warning
57+
58+
On May 24, The Decoder reported that despite the Pentagon flagging Claude as a supply chain risk, Anthropic may continue supplying Claude to the NSA.
59+
60+
This reflects the complex balance AI companies must strike between national security demands and security oversight.
61+
62+
> **Awesome AI View:** A supply chain risk flag is a serious security signal, but the NSA's demand for AI capability is non-negotiable. Anthropic faces a choice: walk away from a major government customer or continue with reputational risk. This also exposes a broader question — when AI models become part of national security infrastructure, how does the definition of "supply chain security" need to evolve?
63+
64+
Source: [The Decoder](https://the-decoder.com/anthropic-may-keep-supplying-claude-to-the-nsa-despite-being-flagged-as-a-supply-chain-risk-by-the-pentagon/)
65+
66+
---
67+
68+
## Microsoft Releases Webwright: Terminal-Native Web Agent Framework
69+
70+
On May 24, Marktechpost reported that Microsoft Research released Webwright, a terminal-native Web Agent framework that scores 60.1% on the Odysseys benchmark, far surpassing base GPT-5.4's 33.5%.
71+
72+
Webwright represents Microsoft's latest exploration in Web Agents — enabling AI to interact with the web directly in terminal environments rather than relying on browser UI.
73+
74+
> **Awesome AI View:** The jump from 33.5% to 60.1% is significant — it demonstrates that specialized agent frameworks are far more effective than simply calling a base model. The terminal-native design philosophy is also interesting: handling web interactions at the terminal layer is more efficient and more programmable than simulating human browser operations. This is a strong move by Microsoft in the AI agent race.
75+
76+
Source: [Marktechpost](https://www.marktechpost.com/2026/05/24/microsoft-research-releases-webwright-a-terminal-native-web-agent-framework-that-scores-60-1-on-odysseys-up-from-base-gpt-5-4s-33-5/)
77+
78+
---
79+
80+
## Microsoft Releases Fara1.5: Browser Computer-Use Agent Family (4B/9B/27B)
81+
82+
On May 22, Marktechpost reported that Microsoft released Fara1.5 (4B/9B/27B parameters), outperforming OpenAI Operator and Gemini 2.5 Computer Use on the Online-Mind2Web benchmark.
83+
84+
Fara1.5 is Microsoft's major positioning in the "Computer Use" direction, offering a complete product matrix from lightweight to high-performance.
85+
86+
> **Awesome AI View:** Three sizes cover different scenarios: 4B for edge deployment, 9B for balancing performance and cost, and 27B for peak capability. Outperforming both OpenAI and Google on multiple benchmarks simultaneously shows that Microsoft's investment in Computer Use has reached a "competitive" stage. This赛道 (track) is heating up fast.
87+
88+
Source: [Marktechpost](https://www.marktechpost.com/2026/05/22/microsoft-releases-fara1-5-a-family-of-browser-computer-use-agents-4b-9b-27b-that-outperform-openai-operator-and-gemini-2-5-computer-use-on-online-mind2web/)
89+
90+
---
91+
92+
## Researchers Let Claude Code Discover AI Scaling Algorithms Humans Probably Wouldn't Have Designed
93+
94+
On May 24, The Decoder reported that researchers let Claude Code autonomously discover AI scaling algorithms that humans probably wouldn't have designed.
95+
96+
This is a meta-scenario: using an AI coding tool to improve AI system training methods — AI designing AI optimization.
97+
98+
> **Awesome AI View:** This is another milestone in AI self-iteration capability. When AI starts designing training algorithms for AI systems, we're one step closer to a "self-improving AI" loop. The key question isn't whether AI can do it, but whether humans can understand and verify these AI-designed algorithms. Interpretability becomes even more critical in this context.
99+
100+
Source: [The Decoder](https://the-decoder.com/researchers-let-claude-code-discover-ai-scaling-algorithms-that-humans-probably-wouldnt-have-designed/)
101+
102+
---
103+
104+
## NVIDIA Releases Gated DeltaNet-2: Decoupling Erase and Write in the Delta Rule
105+
106+
On May 24, Marktechpost reported that NVIDIA AI released Gated DeltaNet-2, a linear attention layer that decouples erase and write operations in the Delta Rule.
107+
108+
This is NVIDIA's latest research advance in efficient sequence modeling, aimed at improving long sequence processing efficiency.
109+
110+
> **Awesome AI View:** Linear attention is one of the most promising alternatives to Transformer-based sequence modeling. Gated DeltaNet-2's core innovation lies in decoupling "memory erase" from "memory write" — similar to how the human brain separates forgetting from learning. If this architecture proves effective at scale, it could provide a new building block for next-generation efficient models.
111+
112+
Source: [Marktechpost](https://www.marktechpost.com/2026/05/24/nvidia-ai-releases-gated-deltanet-2-a-linear-attention-layer-that-decouples-erase-and-write-in-the-delta-rule/)
113+
114+
---
115+
116+
## Google AI Glasses Hands-On: "Almost There"
117+
118+
On May 22, TechCrunch published a hands-on review of Google's AI glasses, concluding that the product is "almost there."
119+
120+
Google's exploration in AI hardware finally has a product form approaching maturity.
121+
122+
> **Awesome AI View:** "Almost there" is an interesting positioning — it means core functionality works, but some key experience elements still need refinement. The success or failure of Google Glasses largely depends on whether the AI features can deliver unique value beyond what a phone provides. If it's just a phone screen on your face, that's not compelling. But if it can differentiate in real-time translation, scene understanding, and environmental awareness, it could be an entirely new interaction paradigm.
123+
124+
Source: [TechCrunch](https://techcrunch.com/2026/05/22/we-tried-googles-ai-glasses-and-theyre-almost-there/)
125+
126+
---
127+
128+
## OpenAI Launches ChatGPT PowerPoint Plugin, Warns It Might Accidentally Delete Your Content
129+
130+
On May 22, The Decoder reported that OpenAI launched a ChatGPT PowerPoint plugin, while simultaneously warning it might accidentally delete user content.
131+
132+
The candid warning itself is news — it highlights that AI generation tools still have significant reliability gaps.
133+
134+
> **Awesome AI View:** OpenAI's honesty is commendable, but it exposes a core problem for AI tools in productivity contexts: users can't trust that AI won't mess up their work. This isn't just a technical problem; it's a trust problem. Before AI tools can truly enter core workflows, "undoability" and "safety guarantees" are prerequisites that must be solved.
135+
136+
Source: [The Decoder](https://the-decoder.com/openai-launches-a-chatgpt-powerpoint-plugin-and-warns-it-might-accidentally-delete-your-content/)
137+
138+
---
139+
140+
## Google CEO Pichai Redefines Search: Links Are Now Just a "Part" of Search
141+
142+
On May 23, The Decoder reported that Google CEO Sundar Pichai now describes links as a "part" of search, redefining the web's role in Google's own product.
143+
144+
This may be one of the most important narrative shifts in Google Search history — moving from "organizing the world's information" to "directly providing answers."
145+
146+
> **Awesome AI View:** When Google starts downplaying the importance of links, it means AI-generated answers are replacing traditional search result pages. The impact on the entire internet ecosystem is profound: website traffic may concentrate further into Google, and content creators' distribution channels become increasingly dependent on AI "understanding" rather than search engine "indexing."
147+
148+
Source: [The Decoder](https://the-decoder.com/google-ceo-pichai-now-calls-links-a-part-of-search-redefining-the-webs-role-in-its-own-product/)
149+
150+
---
151+
152+
## Other Notable Developments
153+
154+
- **Tencent Open-Sources TencentDB Agent Memory**: A 4-tier local memory pipeline providing structured memory capabilities for AI agents. Source: [Marktechpost](https://www.marktechpost.com/2026/05/23/tencent-open-sources-tencentdb-agent-memory-a-4-tier-local-memory-pipeline-for-ai-agents/)
155+
- **Perplexity Open-Sources Bumblebee**: A read-only supply-chain scanner for developer endpoints, designed for security auditing. Source: [Marktechpost](https://www.marktechpost.com/2026/05/23/perplexity-open-sources-bumblebee-a-read-only-supply-chain-scanner-for-developer-endpoints/)
156+
- **Nous Research Releases CNA (Contrastive Neuron Attribution)**: Sparse MLP circuit steering without SAE training or weight modification. Source: [Marktechpost](https://www.marktechpost.com/2026/05/23/nous-research-releases-contrastive-neuron-attribution-cna-sparse-mlp-circuit-steering-without-sae-training-or-weight-modification/)
157+
- **VentureBeat reports Google redesigned its search box for the first time in 25 years**. Source: [VentureBeat](https://venturebeat.com/technology/google-just-redesigned-the-search-box-for-the-first-time-in-25-years-heres-why-it-matters-more-than-you-think)
158+
- **TechCrunch on Spotify's AI bet**: More of everything, less of what you want. Source: [TechCrunch](https://techcrunch.com/2026/05/22/spotifys-ai-bet-more-of-everything-less-of-what-you-want/)
159+
- **TechCrunch reports Ferrari using IBM AI to create F1 superfans**. Source: [TechCrunch](https://techcrunch.com/2026/05/23/ferrari-is-using-ai-to-create-f1-superfans/)
160+
- **TechCrunch reports AI being used to resurrect voices of dead pilots**. Source: [TechCrunch](https://techcrunch.com/2026/05/22/ai-is-being-used-to-resurrect-the-voices-of-dead-pilots/)
161+
- **The Decoder reports a top law school draws a hard line against AI in legal education**. Source: [The Decoder](https://the-decoder.com/one-of-the-worlds-top-law-schools-draws-a-hard-line-against-ai-in-legal-education/)
162+
- **Cloudflare CEO Prince says builders and sellers are safe but AI is coming for the measurers**. Source: [The Decoder](https://the-decoder.com/cloudflare-ceo-prince-says-builders-and-sellers-are-safe-but-ai-is-coming-for-the-measurers/)
163+
164+
---
165+
166+
## One-Line Summary
167+
168+
DeepSeek turned the price war into a permanent strategy, Anthropic's Claude finds bugs faster than humans can fix them, Microsoft is flooding the agent framework space, and Google's AI hardware and search narratives are both undergoing pivotal shifts. AI competition has expanded from "who's smarter" to "who's cheaper" and "who can work autonomously."

0 commit comments

Comments
 (0)