SI Agent News

A delete-button request. A post on community.openai.com asks for a delete button on ChatGPT Work sessions. CONFIRMED: the request exists. STILL OPEN: whether OpenAI builds it. Takeaway: know how to remove client material from any chat tool.

A family is suing OpenAI and CEO Sam Altman over ChatGPT advice. Technews.tw reports that a 31-year-old in Seattle died suddenly after stopping an epilepsy medication on their own. The family points to ChatGPT's advice.

Status: REPORTED, the filing. STILL OPEN: what ChatGPT actually said, whether it caused the outcome, and how a court rules. These are allegations, and the report's headline and summary differ on a detail about the person, so hold specifics loosely.

Analysis: Naming Altman personally is unusual. If the case goes anywhere it could shape how chatbot makers are held responsible, but a filing is not a precedent. The lesson needs no verdict: a chatbot answer is not a prescriber's, lawyer's or accountant's decision.

What this means for you:

• Small business owner: write a one-line rule that health, legal and safety answers need a qualified person's sign-off before anyone acts.
• Job seeker or student: use a chatbot to prepare questions for a doctor, advisor or professor, not to replace them.
• Creator: have an expert review any AI-drafted health content before you publish.

Altman to young founders: stop chasing old trends. At Stanford's Graduate School of Business (a business school), Sam Altman spoke about AI, jobs and entrepreneurship. Technews.tw reports he called the AI era the best time to start a business and said he is 'continually disappointed' that young people chase trends from five years ago.

Status: REPORTED remarks. Opinion, not news, and from the head of a company that sells AI tools.

Analysis: Chasing a five-year-old trend probably means copying what already worked with an AI label on it. Three tests: what can I do now that was impractical two years ago, who pays and how do I know, and what breaks if the model gets worse or the price changes?

For builders: Anue reports that a rival to Nvidia (a chip maker) is entering Oracle's cloud (a cloud provider), with OpenAI reportedly set to buy. The summary does not name the rival. REPORTED, not confirmed. The purchase is STILL OPEN.

Background, not from the report: a model is a large set of learned numbers called weights. For each word it writes, the chip multiplies through them and reads them from memory, so speed and cost depend on both math speed and data movement. Nvidia's edge is also years of tuned software, so a rival must match the code, not just the hardware. Being in a major cloud means customers can rent it instead of installing it.

If true, OpenAI gets a second supplier and price leverage. Confirmation would be Oracle, OpenAI or the chip maker saying so on the record. A deal described only as a trial would count against it. Nothing changes in your API calls today; avoid tying your product to one provider's special features.

The claims check (illustrative). For anyone forwarding chatbot answers to bosses or clients. Price not stated; any chat tool.

Decision this feeds: [what]. If wrong: [consequence]. Using only the material below, list numbered claims labelled QUOTED (copy it), GENERAL KNOWLEDGE or GUESS. Say what you can't know and which role must confirm.
Material and question: [paste]

Check the two riskiest claims. Verify: QUOTED claim not in your text? Tighten labels. Limits: can mislabel confidence; sees only your paste.

Save the claims check as a reusable template. Estimate, not a sourced figure: sorting a confident AI answer into solid and shaky parts by hand might take about 15 minutes; with labels already there, about 5. That is roughly 10 minutes saved per important answer, or about 40 minutes a week at four answers, most of a lunch break.

Store the labelling rules where your tool keeps reusable instructions: ChatGPT's custom instructions, Claude's Projects or Gemini's Gems. Today's stories state no prices, so check each company's pricing page for the plan you need.

What you give up: answers get longer, and the saving only holds if you still do the two-claim check.

Weak: 'Give me a good AI startup idea.' The model doesn't know you, so it returns the most common answer, Altman's old-trend problem.

You are a skeptical startup advisor. My background: [skills]. I can spend [hours/week] and [money]. I can reach [type of customer]. Propose three businesses possible with AI tools but impractical two years ago. For each: who pays, why, how I could talk to ten of them this month, and my first one-week test. Label every assumption known, likely or guess. Say which idea is closest to a crowded trend.

Expect three idea cards. Swap in your industry or zero budget. Before building, find one real payer; demand evidence for 'known'.

Not from today's stories: long-studied human-factors research (how people work with machines); no single paper cited. Automation bias is trusting a machine over your judgment, even against other evidence, especially when answers sound fluent and confident, like chatbot answers.

This week: before acting on a high-stakes AI answer, check one claim against a primary source or ask one qualified person. Write that rule down now; advance rules beat pressured judgment calls.

Still open: OpenAI's lawsuit response and court filings (only family allegations); Oracle, OpenAI or the chip maker confirming the reported purchase; any OpenAI reply on the delete-button request. Analysis: Agentic AI, open-source, and policy were among the busiest areas.

I asked my chatbot to sound more confident. Now it's wrong with excellent posture.

Learning with a reason today: a way to check high-stakes answers, a founder prompt and a time saver, all from today's sources. If this made you smarter, share it with one person who would find it useful. Thank you for reading.