Nội dung
- ChatGPT Work: AI is moving closer to the role of a digital “colleague”
- Long-Horizon Terminal-Bench: AI is still weak at long, multi-step work
- Gemini 3.5 Pro benchmark leak: the foundation model race heats up again
- ChatGPT Work and remote task models: AI is no longer tied to one screen
- The Mitch McConnell image controversy: trust risks when AI enters politics
- Anthropic launches a free Claude course: a clear signal of “AI learns fast”
- When a published image is questioned, the brand message can lose force immediately
- “McConnell returns as AI”: memes, satire, and the new frontier of digital media
- A perspective for the Vietnamese market
- References
This week, AI once again showed two very different sides: on one hand, the rapid commercialization of work tools; on the other, the race for benchmarks, reliability, and the question of whether models can truly handle long, complex tasks. For Vietnamese marketers, this is more than tech news — it is a signal about content production, workflow automation, and information risk management in the AI era.
- Key points:
- OpenAI was mentioned with ChatGPT Work and the ability to assign tasks from mobile/web, showing that AI agents are moving deeper into the workplace.
- The Long-Horizon Terminal-Bench benchmark highlights the challenge of AI completing long tasks, not just “starting well” but also finishing the job.
- Leaks about Gemini 3.5 Pro continue to heat up the foundation model race, but leaked information still needs to be verified through real-world testing.
- Anthropic, along with the viral wave of content around Mitch McConnell, shows that AI is both a productivity tool and a source of controversy over image authenticity.
ChatGPT Work: AI is moving closer to the role of a digital “colleague”
The most notable signal in this week’s batch of news is Greg Brockman’s emphasis that ChatGPT Work is “very good,” along with reports that many Codex capabilities now appear directly in ChatGPT on mobile and web. The standout feature is that users can assign an AI task to run in the cloud straight from their phone, or hand off work already running on a computer to continue on another device.
The significance of this change is not just about the interface or convenience. It reflects a larger trend: AI is no longer just a question-answering tool, but is becoming a layer for coordinating work. For marketing teams, that opens up the possibility of assigning AI tasks such as summarizing documents, drafting copy, classifying campaign data, or continuing unfinished work without being tied to a desk.
Source: Greg Brockman and a quote from @nunezvice on X.
Long-Horizon Terminal-Bench: AI is still weak at long, multi-step work
While AI products are becoming easier to use, the Long-Horizon Terminal-Bench (LHTB) benchmark reminds the market that “doing a few initial steps” is not enough. This benchmark is designed to test whether AI agent can maintain progress through hundreds of interdependent actions, rather than only completing the easiest part of a task.
According to the published information, LHTB includes 46 reproducible terminal tasks across 9 topic groups, evaluating 18 advanced models within the same measurement framework, with durations of up to 90 minutes and around 120–320 steps per task. Preliminary results show the best average reward is only 0.505; no model has solved even one-third of the benchmark, and 29/46 tasks have never been solved by any model.
This is an important reminder for marketers and operators: AI agents are still very strong at initiation, but remain weak at maintaining state, handling unexpected errors, and reliably finishing work. In other words, if AI is to truly replace or support a workflow, the evaluation standard must be closer to “real work” than to a short test.
Source: Yucheng Shi on X about Long-Horizon Terminal-Bench.
Gemini 3.5 Pro benchmark leak: the foundation model race heats up again
Another post drawing attention was a leak about Gemini 3.5 Pro, which reportedly outperforms Claude Fable 5 and GPT-5.6 in internal evaluations, while also showing major improvements over Gemini 3.1 Pro. The post also says Google is preparing for a public release, with a target launch date of 17/7.
Even so, this is still a leak and should be read with caution. In practice, internal benchmarks may paint an optimistic picture, but real-world deployment — from stability and cost to usefulness in specific tasks — is what ultimately determines a model’s position.
For marketers, the story here is that competition among AI platforms is changing the pace of feature launches and the value they deliver. As large models keep upgrading, content, performance, and product marketing teams can expect increasingly powerful tools; at the same time, they must remain skeptical of “outperforming” claims that have not been independently verified.
Source: Entelligence AI on X.
ChatGPT Work and remote task models: AI is no longer tied to one screen
In the fourth item, the message emphasized is that ChatGPT has become more “capable” in recent days, especially thanks to many Codex capabilities appearing directly in the ChatGPT app. Users can tap Work on mobile or web to assign a task that runs in the cloud without sitting in front of a computer, or use Remote to continue work already running on a personal computer.
The valuable point here is the shift from “chatting with AI” to “orchestrating AI to do work.” For marketers, this opens up a more flexible operating model: start a task in the office, monitor progress on a phone, then hand it off or continue on another device. This approach is especially useful for distributed teams that need to react quickly to data, content, or customer feedback.
The broader message is that AI agents are gradually being designed for multi-device, multi-context use. That convenience can boost productivity, but it also requires clearer review and permission processes if businesses apply it to real work.
Source: Greg Brockman and a quote from @nunezvice on X.
The Mitch McConnell image controversy: trust risks when AI enters politics
A series of viral posts continued to revolve around the photo and announcement about Senator Mitch McConnell, as many social media users claimed the published image showed signs of AI or AI editing. In the quoted comments, there were not only doubts about the image’s authenticity, but also debate about his health and ability to continue serving in office.
Although the posts reflect different levels of speculation, the key point for media and marketers is how quickly an image can be scrutinized, doubted, and reinterpreted online. In this information environment, even a single photo without clear context can trigger a trust crisis, especially when it involves politics, health, or public figures.
Source: posts by Doog, tales_typoz, Travis Akers, and PamphletsY on X; along with a quoted C-SPAN announcement about Mitch McConnell.
Anthropic launches a free Claude course: a clear signal of “AI learns fast”
Anthropic also appeared this week with a free 4-hour course that was widely shared on social media. The content mentioned focuses on how to prompt Claude, why Claude can be “less intelligent” when working with code in some situations, how Anthropic uses Claude every day, and some tweaks that help the model perform better.
Even if the wording online is somewhat exaggerated, the core message is quite clear: knowledge about how to work effectively with AI is increasingly being packaged into short, practical, and accessible training content. For marketers, this matters especially because the advantage no longer lies only in “knowing how to use AI,” but in knowing how to build workflows, write prompts, create feedback loops, and embed AI into real work.
If Anthropic or other companies continue to push market education, the gap between casual users and professional users will depend not only on the tools, but also on the ability to learn quickly and apply them correctly.
Source: Ajit kumar on X about the Anthropic course.
When a published image is questioned, the brand message can lose force immediately
In another post on the same McConnell topic, one opinion suggested that the image released by his team had sparked “bipartisan” suspicion that it was AI-generated or contained AI content. While this view comes from social media and does not replace journalistic verification, it shows an important reality: suspicion about an image alone can be enough to dilute both political and media messaging.
For marketers, the lesson lies in verifying communication assets before release. An image seen as untrustworthy can make the entire accompanying statement less convincing. In a context where generative AI can easily create images, videos, and text, brands need to pay more attention to provenance, verification processes, and publication context.
Source: Travis Akers on X.
“McConnell returns as AI”: memes, satire, and the new frontier of digital media
The final item in this roundup has a clearly satirical tone: an X account posted content saying “Senator McConnell announced plan to come back to life as AI”. While this is obviously not serious news, it still shows a very real phenomenon — AI has become a common language for news, commentary, and political memes alike.
From a communications perspective, this reflects the rapid “AI-ification” of public discourse. Any high-attention event can be pulled into an AI story, whether to cast doubt, mock it, or hype it up. For marketers, this is both an opportunity to understand how the public reacts to AI-related content and a reminder to be careful when using AI elements in brand messaging.
Source: PamphletsY on X, along with a quoted Mitch McConnell announcement posted by C-SPAN.
A perspective for the Vietnamese market
For businesses and marketing teams in Vietnam, this set of news points to three very clear priorities. First, treat AI agents as digital staff that need control processes, not as perfect machines; benchmarks like LHTB remind us that long, multi-step work is still a weakness. Second, the trend toward ChatGPT Work and multi-device models signals a new way of working: assign tasks, monitor them, and continue them anywhere.
Third, the McConnell image controversy is a direct warning for brand communications. In an environment where AI can create highly realistic content, Vietnamese businesses need standards for verifying assets, clear source attribution, and a response process for when content is questioned. In short: AI is opening up new productivity, but trust is what will determine who can benefit from it in the long run.
See more marketing news and guides at https://marketing365.vn.
Follow more updates from Marketing365 to stay on top of the latest marketing trends.
Read more articles in the AI Updates category.
This article focuses on latest AI news with a perspective for the Vietnamese market.
References
- Mitch McConnell may or may not be alive…
- Long-Horizon Terminal-Bench
- Gemini 3.5 Pro benchmark leak
- ChatGPT Work
- McConnell photo skepticism
- Anthropic free course
- McConnell image speculation
- McConnell AI comeback joke
