Where Is Artificial Intelligence Headed? The Latest Changes, Controversies, and Breakthroughs (AI News August 2025)
The development of artificial intelligence is not slowing down. Groundbreaking updates, ethical dilemmas, and the growing presence of AI in everyday life show that we are facing changes comparable to the Industrial Revolution. In August 2025, educational innovations, advanced agents, disturbing comparisons to the Manhattan Project, and attempts to win over researchers’ minds are taking center stage.
Table of contents
-
AI Agents: artificial intelligence passes CAPTCHA tests. How “human” is it already?
-
Apple’s “killer quarter”: the company is buying time for its AI strategy
-
Ads in conversations with Alexa? Amazon CEO’s controversial vision
-
Google AI introduces TTD‑DR – Test‑Time Diffusion Deep Researcher
.
Practical Developments in Artificial Intelligence. August 2025.
New feature in ChatGPT: study mode.
OpenAI has launched a new learning mode in ChatGPT that aims to become a fully-fledged digital tutor. Unlike the classic question-and-answer mode, the new feature guides users through the material step by step: it asks guiding questions, adjusts the level of difficulty, and encourages reflection. The creators collaborated with educators, applying principles such as metacognition and cognitive load management.
The goal is not to provide answers quickly, but to learn how to think. This is a response to the growing problem of academic fraud—nearly 7,000 cases of AI-related misconduct were recorded in the UK alone last year. OpenAI acknowledges that this will not solve the problem entirely; a change in the approach to assessment in education is also essential.
Student Maggie Wang admitted that after a three-hour session with this mode, she finally understood sinusoidal positional encoding—a topic that had previously been completely incomprehensible to her. She compared it to a tireless, patient teacher.
.
New Agent Mode in ChatGPT
The new Agent mode in ChatGPT is a significant step toward genuine artificial intelligence autonomy. OpenAI made this feature available in July 2025 to users on the Pro, Plus, and Team plans. In practice, this means that ChatGPT not only answers questions but can also independently carry out complex tasks—from browsing the internet and filling out forms to preparing presentations and analyzing data. Agent mode is not a separate model, but an extension of existing models (such as GPT‑4o) that gain access to tools, a browser, and a virtual computer. The entire process runs in real time, allowing users to track progress on screen.
.
This mode works like a virtual assistant. The user assigns a task in natural language, for example, “plan a trip to Lisbon and find three hotels with a rating above 8,” and the Agent gets to work: opening websites, analyzing data, comparing options, and then presenting the result, often in the form of a file or table. If the task requires a decision or action that could have consequences (such as making a purchase or sending an email), the Agent pauses and asks the user for confirmation. It can be stopped, taken over, or given modified instructions at any time.
The Agent feature also integrates with popular services such as Gmail, Google Drive, and GitHub, enabling you to create documents, analyze spreadsheets, or manage projects directly from ChatGPT. The possibilities are extensive—you can automate everyday tasks, create personalized reports and schedules, and even generate shopping lists. In tests, the agent was able, for example, to prepare a weekly meal plan, generate a shopping list, compare store offers, and present everything in a clear table.
The simplest application: web scraping and gathering information from many useful sources. When it comes to creating attractive reports, agent mode does not perform particularly well.
.
Creating Consistent Characters from a Single Photo
Startup Ideogram has unveiled a tool that can generate consistent characters based on a single image—preserving the face, hairstyle, and angle of the lighting. It is useful, among other things, for creators of comics, games, and avatars. This solution brings us closer to the era of hyper-personalized graphics.
Microsoft Edge with Copilot
The new Copilot mode in the Edge browser analyzes all open tabs and creates summaries and comparisons. It enables voice control and features a unified interaction bar. This is another step toward an “assistance-enabled browser”—a workspace supported by AI.
Google AI Search and Live Mobile Camera
Google is expanding its AI Search, enabling image uploads and soon PDF files. The new Canvas tool allows users to plan tasks visually. On mobile devices, live search using the camera is available—an actual step toward recognizing the surrounding world in real time.
Nvidia Llama Neotron Super 1.5 – benchmark leader
The new version of the Llama Neotron Super 1.5 model achieves record-breaking results in benchmarks for mathematics, reasoning, coding, and chat. Importantly, it runs efficiently on a single Nvidia H100 GPU. This could potentially be a huge change for companies that want to develop their own models locally, without needing to use gigantic supercomputers. Although a computer using graphics cards costing over PLN 100,000 is not a cheap solution, medium-sized and larger companies may be able to afford it.
."""
Adobe Photoshop – AI that helps modify photos.
Photoshop has received three significant AI features:
- Harmonize – automatic matching of the lighting and tones of objects inserted into an image;
- Generative Upscale – improving photo quality up to 8 megapixels;
- Improved object removal – now more effective and faster.
AI news
AI Agents: artificial intelligence passes CAPTCHA tests. How “human” is it already?
In an experiment involving ChatGPT agents embedded in a secure testing environment with internet access, the AI managed to pass the CAPTCHA protecting websites using Cloudflare Turnstile. The agents clicked the “Verify that you are human” box, and the system did not challenge their identity—it did not initiate any additional image-based tasks.
Turnstyle analyzes subtle signals: the way the mouse is moved, the browser fingerprint, and IP history. The AI agents behaved naturally enough not to arouse suspicion. What’s more, they later successfully performed complex tasks, such as ordering groceries.
It’s not just a technological breakthrough; it’s a signal that the line between “human” and “bot” is beginning to blur.
Sam Altman: “Testing GPT-5 Is Like the Manhattan Project”
OpenAI CEO Sam Altman compared testing the GPT-5 model to work on nuclear weapons. According to him, GPT-5 does not merely respond faster—it understands. He points to the models’ growing “contextual awareness” and the need for a radical increase in oversight.
Altman laments that “there are no adults in the room” – regulatory institutions are failing to keep pace with the development of AI. His statement is a warning: we are entering territory that may bring not only benefits, but also real risks to society.
On the other hand, Sam Altman is known for exaggerating the significance and capabilities of current artificial intelligence models.
1 Billion Dollars? Not Enough. Meta Hits a Value Ceiling
Mark Zuckerberg and Meta Labs wanted to lure top AI researchers to their side, offering them as much as one billion dollars over several years. Among those targeted were members of Mery Morati’s team at Thinking Machines Lab. They all turned down the offer.
This shows that in the world of AI researchers, money has ceased to be the only currency. Trust, values, and scientific autonomy are more important. At a time when technology is entering the realm of ethics, personnel decisions take on deeper significance.
Apple’s “killer quarter” buys the company time for its AI strategy
Apple published its report for the third fiscal quarter of 2025, beating analysts’ expectations: revenue exceeded USD 94 billion, driven by a 13% increase in iPhone sales (USD 44.6 billion), while services reached a record USD 27.4 billion. The long-term stock price fell by only 2–3% after the release, signaling that the market viewed this stability favorably, albeit with reservations about further AI investments. CEO Tim Cook announced significant spending on artificial intelligence, as well as potential acquisitions: the startup Perplexity AI. is reportedly under consideration. Analysts warn, however: Apple still remains behind Nvidia and Microsoft in key AI solutions, including the delayed rollout of its new LLM-based Siri. They point out that although the company has delivered strong quarterly results, its future success depends on the speed of AI integration and the clarity of its strategy.
Ads in Conversations with Alexa? Amazon CEO’s Controversial Vision
TechCrunch reports (dated July 31, 2025) that Amazon’s CEO has begun preparations to introduce advertisements directly into conversations with Alexa, in response to voice commands generated by users. This would apply in situations where the system provides an answer or suggestion: at these moments, contextually tailored ads could appear, such as “Nearest ice cream shop: via a sponsored message.” The company argues that this is a natural extension of the service’s monetization model; however, experts point to the risk of degrading the user experience and eroding trust, particularly regarding privacy and the use of voice data.
Google AI Introduces TTD‑DR – Test‑Time Diffusion Deep Researcher
The Google Cloud AI Research team (including Rujun Han, Yanfei Chen, and others) has published the framework for a new TTD‑DR framework, which models the creation of a research report as a test-time diffusion process. The system starts with a “rough” draft of the proposal and then iteratively refines it by performing a search for missing information (retrieval), analyzing the gaps (gap analysis), and subsequently improving the draft. The entire pipeline optimizes itself through a self-evolution algorithm that generates, evaluates, and combines better variants of research plans, queries, responses, and the final text. Compared with other DRA agents (e.g., OpenAI Deep Research), TTD‑DR achieves at least statistically significantly better results on complex benchmarks requiring multi-step reasoning.
Does the name refer to the famous abbreviation TL;DR (too long, didn’t read)? I don’t know, but I suspect so.
What does it all mean?
The innovations introduced by OpenAI, Google, Microsoft, Nvidia, Adobe, and Ideogram share one common denominator: the shift of AI from a tool to a collaborator. AI is no longer merely a support. Increasingly, it is becoming an entity—a consultant, teacher, and negotiator.
On the other hand, there is a growing need for reflection and regulation. Altman refers to the Manhattan Project for good reason—GPT-5 and its descendants could affect society as profoundly as the invention of the atomic bomb.
At the same time, we are witnessing a return to values in science—money is no longer the only argument. Researchers want to influence the direction of technological development, not merely count profits.
Artificial intelligence is no longer something that “might change the world someday.” It is already changing it—every day, quietly, but irreversibly.
As users and creators, we must learn not only how to use AI, but also understand what it means to be human in an age when a machine can pass a CAPTCHA test.
Sources: