The return of open-weight models and cracks in trust in AI

Meta is back to open source, and Kakao is splitting off an AI company. Meanwhile, more than half of Americans say AI still makes them uneasy. Today's briefing looks at the gap between 'technology diffusion' and 'social acceptance'.

---

1. Meta releases 30B open-weight model 'Muse Glimmer'

• Source: TechCrunch / VentureBeat | 2026.08.10 | Link
• What happened: Meta is reported to have released 'Muse Glimmer', a 30-billion-parameter open-weight model under the Apache 2.0 license. It is designed to run agent workloads (AI that carries out tasks autonomously) locally on a Mac or PC with a single consumer graphics card. Timed with the release, CEO Mark Zuckerberg argued that US regulatory barriers to open-source AI should be lowered, and said the weights of the higher-end 'Muse Spark 1.2' would be released as well.
• Why it matters: It is a step back from the trend of locking frontier models behind APIs. A 30B-class agent model that can run locally greatly reduces the data exfiltration burden for companies.

2. Kakao's board approves a spin-off into 'Kakao AI' and 'Kakao X'

• Source: ZDNet Korea | 2026.08.21 | Link
• What happened: Kakao's board is reported to have approved a spin-off dividing the company in two. The AI business built around KakaoTalk goes to the new entity 'Kakao AI', while future businesses based on core subsidiaries in techfin, content, and mobility go to the surviving entity 'Kakao X'. Kakao had earlier partnered with Coupang Eats to deliver the first agentic AI that understands conversational context and completes everything from authentication to ordering in a single flow.
• Why it matters: Among Korean platform companies, it is a first-of-its-kind decision to carve AI out into a separate legal entity. It reads as a move aimed at both raising investment capital and speeding up the business.

3. OpenAI accelerates its pursuit of Anthropic in the enterprise market

• Source: TechCrunch | 2026.08.20 | Link
• What happened: New data reportedly shows OpenAI narrowing the gap with Anthropic among business users. The enterprise API market has until now been considered an area of Anthropic strength. Meanwhile, a tally showed Anthropic recorded Q2 2026 revenue of $10.9 billion and its first operating profit of $559 million.
• Why it matters: At the enterprise adoption stage, price, governance, and distribution channels decide the outcome as much as model performance. Korean companies' vendor selection criteria are likely to move along with it.

4. "AI hasn't won people over yet" — 52% more concerned than excited

• Source: TechCrunch | 2026.08.19 | Link
• What happened: A Pew Research Center survey was cited in which 52% of Americans said they are 'more concerned than excited' about the growing role of AI in daily life, a sharp rise from 37% in 2021. The analysis is that usage is rising but acceptance is not keeping pace.
• Why it matters: It signals that in content and marketing, 'AI use disclosure' and transparency are becoming the default rather than an option.

5. OpenAI expands cyber defense service 'Daybreak'

• Source: TechCrunch | 2026.08.10 | Link
• What happened: With AI-powered attacks on the rise, OpenAI is reported to have announced an expansion of 'Daybreak', the cyber defense service it launched earlier this year. Anthropic had previously released its security-focused model 'Mythos'.
• Why it matters: As both offense and defense move to AI, security is establishing itself as a new revenue category for AI labs.

6. Adobe Firefly's unified generate-and-edit workspace beta

• Source: Adobe / Creative Bloq | 2026.08 | Link
• What happened: Firefly released a beta of a unified workspace that brings image and video generation and editing, project media, generation history, reusable 'elements', and a video timeline together on one screen. A batch run (Run) interface that processes up to 500 input assets at once was added as well. It currently runs on Firefly Image Model 4 and the premium 4 Ultra, and Image Model 3 was fully retired from the service this month.
• Why it matters (AEM and enterprise perspective): Batch generation in units of 500 targets enterprise DAM (digital asset management) scenarios rather than individual creation. Combined with AEM Assets, automating the pipeline from mass production of campaign variants to asset registration to distribution becomes a realistic option. The retirement of Model 3 leaves an issue of managing look-and-feel consistency for existing generated assets.

7. AI agents deployed per company roughly triple in a year

• Source: Salesforce Agentic Enterprise Index (as cited) | 2026.08 | Link
• What happened: Per Salesforce's tally, the number of AI agents deployed per organization was cited as rising from an average of 5 in early 2025 to 13 in April 2026. Agent build time fell 53%, and employee agent usage sessions tripled.
• Why it matters: The discussion has moved from 'whether to adopt AI' to 'how many to build and how fast'. Demand for agent operations and management (governance) is the next bottleneck.

---

Today's insight

1. Openness and localization are regaining momentum. Meta's 30B open-weight release takes direct aim at the corporate constraint of 'we can't send data outside'. The more regulated the industry, the greater the value of evaluating local agents.
2. The pace of technology diffusion and the pace of social acceptance are diverging. Agents per company tripled, yet public concern rose from 37% to 52%. This gap is likely to be filled by regulation and disclosure requirements.
3. The front line of AI competition is moving from 'models' to 'security and operations'. OpenAI's Daybreak expansion, Anthropic's Mythos, and Adobe's batch pipeline are all aimed at the operating floor, not benchmark scores.