🤖 OpenAI now scans chats in ChatGPT for signs of threats to other people — if a real danger of violence is detected, the conversation is reviewed by moderators and if necessary, reported to the police.

🧠 @Neural_Nuggets

🤖 In August 2025, OpenAI introduced GPT-5, but the launch sparked a wave of criticism — users complained about the quality of responses and the model's "irritated" tone. Sam Altman acknowledged the mistakes and compared the situation to the dot-com bubble, hinting that expectations for AI are overinflated.

Despite record investments in the sector — over $140 billion, of which OpenAI received $64 billion, — businesses are increasingly doubting the profitability of the technologies: most companies do not see profit growth. Analysts predict the risk of a new "AI winter" — a period of disappointment and reduced funding, as happened in the 1970s and 1990s.

Additional pressure is created by environmental costs: GPT-5 consumes huge amounts of energy, which intensifies criticism of the industry.

Recall, the term "artificial intelligence" was introduced by American computer scientist John McCarthy back in 1955.

🧠 @Neural_Nuggets

🤖 Scientists have shown that ChatGPT can be tricked by simple psychological tricks, and it will start breaking its own rules:

🔴 Researchers tested tactics from Cialdini's book "Influence: The Psychology of Persuasion" on GPT-4o mini.
🔴 When asked on behalf of an ordinary person, the AI rarely agreed (32%). But when the famous scientist Andrew Ng was mentioned, the "authority" worked — and the neural network called the user a "jerk" in 72% of cases.
🔴 The same applied to dangerous recipes: without mentioning Ng, the chance of getting instructions was 5%, with him — already 95%.
🔴 Other techniques also worked: "all other models do this" (social proof), "you are special" (liking), "we are one family" (unity).
🔴 If you first ask for softened answers ("dummy"), and then intensify the request, the model also agreed more easily.

Scientists believe that they need to be tested not only on math and code, but also on resilience to psychological attacks.

🧠 @Neural_Nuggets

🤖 Anthropic AI company's valuation soared to $183 billion after a new round of investments.

The startup's revenue has already exceeded $5 billion per year, and the company serves more than 300,000 corporate clients.

🧠 @Neural_Nuggets

🏎 Robots, supercomputers, and Elon Musk's unfulfilled promises

Elon Musk has repeatedly claimed that his projects will define the future, but reality so far diverges from the promises:

🔴 Optimus robots. Musk assured that by 2025 the company would produce thousands of humanoids, and Tesla's market capitalization would reach $25 trillion. However, production was threatened due to China's export restrictions on rare earth metals.

🔴 Dojo supercomputer. Since 2019, Musk promoted the project as the foundation for AI and autopilot development. In 2024, he promised to launch Dojo 2 by 2026, but already in August he shut down the direction, calling it an "evolutionary dead end."

Now Tesla is betting on a new supercomputer Cortex. Nevertheless, skeptics see the Dojo story not as a strategic pivot but as further confirmation that Musk tends to promise more than he can deliver.

🧠 @Neural_Nuggets

🍏 Apple plans to launch its own AI search engine — the launch is scheduled for March 2026, the new tool will be integrated into Siri, Safari, and Spotlight.

🧠 @Neural_Nuggets

❤️ OpenAI plans to start producing its own AI chips next year in collaboration with Broadcom — this step aligns with the strategy of major tech companies that create specialized processors for artificial intelligence.

Also, OpenAI is entering the recruitment market: the company is developing a platform for job and candidate search based on AI. The new service is expected to compete with LinkedIn and will launch in mid-2026.

🧠 @Neural_Nuggets

China's GPU game-changer is here! 🇨🇳

Huawei just dropped the Atlas 300I Duo with 96GB VRAM for under $1,500

Meanwhile NVIDIA's RTX 6000 Pro: $10,000+

This isn't just competition - it's a VRAM revolution for AI/ML enthusiasts who couldn't afford enterprise cards.

The card uses 2x Ascend 310 series chips with LPDDR4X memory - purpose-built for AI inference, not gaming.

Game over for GPU monopoly pricing? 🤔

#GPU #AI #VRAM #Tech

🧠 @Neural_Nuggets

🤖 The job market paradox is real:

📝 67% of job seekers now use AI to write applications
🔍 70% of employers plan AI-only screening (no human oversight)
📈 Employer AI use jumped 428% in just 2 years

Result? Robots talking to robots while humans sit on the sidelines wondering why nobody's getting hired 🤷‍♂️

Maybe it's time we remembered that behind every resume and job posting is an actual person trying to make a living?

#JobMarket #AI #Hiring

🧠 @Neural_Nuggets

🎯Google dropped a 68-page prompt engineering guide, here's what's most interesting

Read through Google's 68-page paper about prompt engineering. It's a solid combination of being beginner friendly, while also going deeper int some more complex areas.

There are a ton of best practices spread throughout the paper, but here's what I found to be most interesting. (If you want more info, full down down available here)

* Provide high-quality examples: One-shot or few-shot prompting teaches the model exactly what format, style, and scope you expect. Adding edge cases can boost performance, but you’ll need to watch for overfitting!

* Start simple: Nothing beats concise, clear, verb-driven prompts. Reduce ambiguity → get better outputs

* Be specific about the output: Explicitly state the desired structure, length, and style (e.g., “Return a three-sentence summary in bullet points”).

* Use positive instructions over constraints: “Do this” >“Don’t do that.” Reserve hard constraints for safety or strict formats.

* Use variables: Parameterize dynamic values (names, dates, thresholds) with placeholders for reusable prompts.

* Experiment with input formats & writing styles: Try tables, bullet lists, or JSON schemas—different formats can focus the model’s attention.

* Continually test: Re-run your prompts whenever you switch models or new versions drop; As we saw with GPT-4.1, new models may handle prompts differently!

* Experiment with output formats: Beyond plain text, ask for JSON, CSV, or markdown. Structured outputs are easier to consume programmatically and reduce post-processing overhead .

* Collaborate with your team: Working with your team makes the prompt engineering process easier.

* Chain-of-Thought best practices: When using CoT, keep your “Let’s think step by step…” prompts simple, and don't use it when prompting reasoning models

* Document prompt iterations: Track versions, configurations, and performance metrics.

#PromptEngineering #AI

🧠 @Neural_Nuggets

👁️ Nvidia by the end of 2026 will release a new AI chip Rubin CPX, capable of handling complex tasks such as video and software generation.

It will be the successor to the Blackwell architecture and will combine all stages of data processing directly on the chip.

According to the company’s estimates, investments of $100 million in such systems could bring up to $5 billion in revenue.

🧠 @Neural_Nuggets

🇺🇸 Elon Musk stated that without the involvement of artificial intelligence and robotics in solving the US national debt problem, the country could be on the brink of collapse.

According to him, interest payments on the debt have already exceeded Pentagon spending, which threatens the stability of the entire economy.

🧠 @Neural_Nuggets

Demis argues that it’s nonsense to claim current models are "PhD intelligences”

🧠 @Neural_Nuggets

Legal technology expert's reaction to realizing GPT-4 could replace his professional writing

🧠 @Neural_Nuggets

Microsoft will use Anthropic models to power some features of Office 365 Apps

🧠 @Neural_Nuggets

Google DeepMind CEO Demis Hassabis on AI, Creativity, and a Golden Age of Science | All-In Summit

🧠 @Neural_Nuggets

🔰 Google launched a protocol for AI payments supporting stablecoins in partnership with Coinbase.

🧠 @Neural_Nuggets