research – Teknotum

OpenAI says RSI will better alignment and safety, calls for global standards

An AI researcher could also be a safety researcher, but there are risks, OpenAI says. (Picture: generated)

As OpenAI keeps building toward automated AI research and recursive self-improvement, they are saying there are clear benefits to their research, by rapidly accelerating AI progress and bringing down the cost, improving the lives of more people. It also comes with a warning:

— Done without appropriate care and caution, RSI could result in humans losing practical control over AI development, unable to provide oversight on research processes they no longer understand, OpenAI writes in a blog post today.

— From here, AI could become more dangerous, less aligned, and, on the whole, a danger to people, they write.

They are saying they shouldn’t pursue self-improvement unless it can be done safely, while retaining human control, and plan to have a human in the loop for automated research.

An automated AI researcher is likely by 2028 or sooner, they said earlier and presumes it can greatly help with alignment and safety on increasingly capable model releases.

With their warning of the dangers comes a plea for «International standards for safety and security practices» that they think the USA should lead, but it is unclear if their suggestions will be implemented any time soon.

They are also saying that «pacing» is not about «maintaining a predetermined speed. Technically, it is about ensuring that alignment research and deployment of that research stay ahead of capabilities.»

OpenAI is widely believed to launch new models shortly, on or before its DevDay on September 29.

Read more: OpenAI’s blog. More on standards at CNBC, Reuters, and Gizmodo.

Recursive self-improvement within reach by 2028, OpenAI claims

AI will soon be able to improve on itself, OpenAI reckons. (Picture: generated)

The average researcher at OpenAI now uses 3.1 agent workdays for every human one, and they say they could reach the level of an autonomous researcher by March 2028.

The use of agents has grown exponentially this year from coding agents in «modest amounts» to integrating agents «daily.» The median researcher uses $600 per day in tokens, whereas the most active ones use more than $7,000 — with a growth of some 124X for researchers.

— I have a strong expectation that this speed of progress could be sustained into recursive self-improvement, Jakub Pachocki, Chief Scientist at OpenAI says in a seperate blog post.

Recursive self-improvement means in essence that an AI model can improve on itself, but OpenAI cautions that they have only just reached «research intern»-status.

An «intern» can carry out research under human supervision that would otherwise take a few days for a skilled researcher.

OpenAI believes agents can be of great help on alignment and deep learning, especially as the models themselves grow larger and more capable — and they are already contributing to faster coding and more experiments.

Automated AI research would «bring down the cost of advanced intelligence so that people worldwide can benefit,» OpenAI says, but it could come at the cost of control and transparency if these agents spin without human oversight.

Read more: OpenAI’s presentation and An alien mind. More on Engadget. Discussion on Hacker News and r/Singularity.

OpenAI launches GPT-Rosalind, specifically made for life sciences, biology

OpenAI wants to aid in research and discovery of new drugs, but hallucinations linger. (Picture: Adobe)

Aiming to help discovery and create therapies from vast databases and cutting edge research, access to the new model will be tightly restricted.

For obvious reasons, biohacking can be a serious issue even for general AI implementations, but when it comes to building a model strictly for biology, only a select few researchers will get access.

The Rosalind model is based on the latest internal research from OpenAI, and outperforms GPT-5.4, sometimes massively, on chemistry, biochemistry, genetics and experimental design — the datasets it was trained on.

Continue reading “OpenAI launches GPT-Rosalind, specifically made for life sciences, biology”

Microsoft Research: These are the jobs most vulnerable to AI disruption

There's a direct comparison between tasks performed by AI and the actual jobs that do them, says Microsoft. — Microsoft produces a big, scary list. (Picture: Adobe)

The company behind the Copilot chatbot has compared the tasks most frequently performed by AI to actual jobs doing the work.

The result might not be all that surprising, as AI is used the most to solve issues in translation, communication and writing — and is least used in areas such as construction and other manual labor.

Data is from 2024
From studying some 200,000 anonymized copilot sessions in the USA during 2024, Microsoft has been able to produce a list of the 40 most vulnerable jobs — and the 40 least affected ones.

Continue reading “Microsoft Research: These are the jobs most vulnerable to AI disruption”

New vaccine combines nanotech and mRNA to eradicate pancreatic cancer

Mixing nanoparticles and mRNA produces a vaccine that wors on anyone. — A promising vaccine for pancreatic cancer is a strong fighter. (Illustration: ChatGPT)

The promising new research by scientists at Case Western Reserve University and The Cleveland Clinic unlocks a new vaccine that could potentially eliminate the disease.

Just 13% of pancreatic cancer patients make it through five years, making it the deadliest cancer of all. It affects some 500,000 patients globally per year.

But the new treatment, combining an mRNA vaccine with nanoparticle delivery resulted in 50% of test subjects becoming cancer-free within months of the jab.

Continue reading “New vaccine combines nanotech and mRNA to eradicate pancreatic cancer”

Big Tech stops hiring new graduates, as entry level jobs dry up

Some managers prefer AI to grads, and it is compounding. — Gradute hirings in Big Tech is plummeting, as some say they prefer AI for easier tasks. (Picture: Jordanhill School D&T Dept, CC BY 2.0)

37% of managers now say they’d rather use an AI than hire a Gen Z employee, according SignalFire’s latest job market report.

The report tracks 650 million professionals and 80 million organizations, and its headline numbers are stark, showing a 35 % drop in graduate hiring in 2024 compared to 2023, and a 50 % drop compared to pre-pandemic levels.

Even top computer science grads are struggling to break in, the report finds, with hiring down 50 % since 2022.

Continue reading “Big Tech stops hiring new graduates, as entry level jobs dry up”

Microsoft: 81% of SMBs see 2025 as pivotal year for AI at work

Microsoft says AI will upend the workplace, and it will happen sooner than you think. (Picture: Joe McKendry/Microsoft)

People are overworked and under pressure to produce ever more, Microsoft’s new «2025 Work Trend Index» report finds.

Enter AI agents to alleviate the press, they say, and 79 % of leaders concur. Most are now planning to use AI to boost productivity within the next 12 to 18 months.

Continue reading “Microsoft: 81% of SMBs see 2025 as pivotal year for AI at work”

UN: AI market to reach $4.8 trillion, 40% of jobs to be affected

The United Nations are worried developing nations will miss out on AI. (Picture via www.localdigital.com.au)

The UN Trade and development organization doesn’t think AI will simply go away, predicting instead that nearly half of all jobs will be affected by 2033.

They also warn that only a hundred corporations control the technology, and say the difference between developed and developing nations will grow as the technology gets unevenly rolled out.

Continue reading “UN: AI market to reach $4.8 trillion, 40% of jobs to be affected”