In partnership with
Sponsor:
Elon's new company is private. These 3 tickers aren't.
The next Apple may already exist. Insider sources say Elon has spent two years building a secret device inside Tesla's facilities — one he claims will be "10x bigger than the largest product in history."
There's just one problem: the company is private, and unless you know Elon personally, you can't buy a single share. That was true until Guardian's research team found three public ticker symbols sitting in the launch supply chain.
Click here to see all 3 tickers, free of charge.
You won't hear these names on CNBC — Wall Street hasn't published a word on the connection. But when the launch hits September 21, that quiet ends.
Some are already calling this the biggest opportunity since AI. For anyone who missed Apple before the iPhone, this may be a second look at that kind of setup.
iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #156 WEDNESDAY · 7 OCTOBER 2026
PREVIEW
THE HOOK
On Monday, New York City Council Speaker Julie Menin asked OpenAI, Anthropic, Google and Meta, all testifying under oath, to assure the public that their agents would always follow the safeguards meant to prevent catastrophic harm. None would promise it. OpenAI said it couldn’t ‘commit or guarantee that any technology is without risk’. Anthropic called the science ‘fundamentally hard and unsettled’. Honest answers. So when an agent’s output becomes a decision, who answers for it? Usually, you.
OUR ANGLE
Nobody can guarantee it. Someone still checks
Six days before the hearing, leaders of six AI and chip companies signed a White House accord promising internal controls, external audits and board oversight. It’s voluntary, sets no penalties and doesn’t require anyone to publish an auditor’s findings. At the hearing, Meta guaranteed only its commitment to safety; Google said promising perfection wasn’t possible. Refusing an impossible promise is the honest answer, not a scandal.
Our read: nobody can guarantee an AI’s output, and a reviewer can’t either. What can be assigned is the duty to check it before it becomes a decision, and California has started writing that into law. SB 947, signed the day after the accord, bars employers from July 2027 from relying solely on an automated system to discipline or fire someone. Where they rely on one primarily, a person must corroborate the decision using the data behind the output or other relevant information. Output that can’t be corroborated, or proves inaccurate, incomplete or misleading, can’t be used. Borrow that standard for any decision you can’t undo. Checking won’t make it certain. It can catch the claim that would have sunk it.
Prediction: by 1 July 2027, when California’s rule takes effect, no more than one of the accord’s six signatories will have published an external auditor’s findings under it. A company’s own summary won’t count.
COMPANION DEEP DIVE
One decision, checked end to end: a worked example where the misleading claim was a quote, a one-line record and a monthly spot check you can repeat.
AI NEWS ROUNDUP
Twelve points, one word in ten and 88%
1 Twelve points. Pew had an AI model answer three of its 2026 surveys as ‘digital twins’ of real panel members. Across nearly 300 questions, the share giving each answer was off by 12.4 percentage points on average. The twins said 98% of Americans knew what the First Amendment guarantees; 52% did. Offered ‘not sure’, real people chose it about four times as often. The move: ask anyone selling synthetic customers what they were tested against.
2 One word in ten. Over the coming weeks, eligible ChatGPT and Codex text in the EU will carry textGrain, an invisible watermark meant to meet the AI Act’s transparency rules; API customers anywhere can opt in on selected models. In OpenAI’s own tests on 400-token passages, swapping one word in ten for a synonym cut detection from about 92% to 66%. And a watermark records where text came from, not whether it’s true.
3 Up to 88%. Reuters has reviewed research on agents built on Chinese models. In one study, agents running Qwen3, Kimi-K2 and DeepSeek-V3.2 made at least one false statement in 84–88% of rounds of a simulated contract-bidding game. US models tested the same way did much the same. The move: when an agent says a task is done, check the result, not the report.
THE THREE SPECIALS / DO · USE · UNDERSTAND
PROMPT OF THE WEEK
A corroboration sheet
Use it before you act on an AI answer that matters. It doesn’t check anything. It shows you what to check.
I’m about to act on the AI output below. Help me review it, not defend it. Don’t rewrite it.
OUTPUT: [the answer, summary or recommendation]
MY SOURCES: [paste the documents or data it should rest on]
DECISION: [what I’ll do, and what it costs if it’s wrong]
1. List only the claims the decision depends on, usually no more than five.
2. For each, say where it comes from: my sources (quote the passage, under 25 words), general knowledge or inference. Only say ‘my sources’ if you can quote it. ‘Not sure’ is an acceptable answer; a confident guess isn’t.
3. For each claim, say what I should open to check it. For quoted claims, name the surrounding text to read too: a quote can be accurate and still misleading.
Why it works: step two tells you where to look, not what to skip. Every claim the decision rests on gets checked, quoted ones included.
TOOL OF THE WEEK
Gemini Notebook
Google’s research tool, called NotebookLM until July. Upload the documents a decision rests on and ask questions; each answer carries inline citations that take you to the passage behind it.
Use it to find evidence, not to summarise. For each claim the sheet lists, click the citation and read the paragraph around it. A citation shows where a claim came from, not whether the model read it right.
Two limits. It only knows what you upload, so a clause you left out stays missing. And your account type matters for confidential files. On personal accounts, Google says uploads aren’t used to train Gemini Notebook unless you send feedback. On Workspace and Education accounts, nothing is seen by human reviewers or used for training.
TIP OF THE WEEK
Make ‘not sure’ a respectable answer
This won’t make a model right. It makes it likelier to tell you when it’s guessing. Add one line wherever accuracy matters: ‘If my sources don’t settle it, answer NOT SURE and say what’s missing.’ OpenAI’s researchers argued last year that most benchmarks reward a guess over ‘I don’t know’, so models learn to guess. A prompt can’t undo that training, but it can tell the model a gap is welcome.
The limit: this doesn’t make a model’s confidence reliable. Pew’s twins were wrong without hedging, and an abstain option only catches the gaps a model notices.
Pro move: slip in one question your sources can’t answer. If it doesn’t come back NOT SURE, don’t trust the ‘sure’ answers around it.
YOUR MOVE
Check one decision before you sign it
Pick one decision this week where AI did part of the thinking: a supplier comparison, a shortlist, a reply to an unhappy customer. Run the corroboration sheet, open the evidence for each claim it lists, and note what held.
REPLY WITH ONE FINDING
Supplier comparison: 4 claims mattered; 3 held. 1 quote left out the notice period, so we switched.
Illustrative reply only. Leave out names and confidential details. I read every reply; we won’t publish a tally.
Someone else in your company signs off AI-assisted work? Forward them this.
Disclosure: the original draft used Claude. Anthropic testified at the hearing and signed the accord, and Pew’s study used Claude Opus 4.6.
P.S. The companion adds the step most people skip: a monthly spot check of the AI tool itself, against a case whose facts you’ve verified.
PUBLISHED BY FRONTWAVE MEDIA LTD · IPROMPT.COM
iPROMPT / 156 /
Hiring abroad? How many Euros could it cost you?
Bringing on someone new to the team? Their salary is just the starting point.
Taxes, benefits, employer contributions, and compliance costs can add up quickly, and vary by country.
Use Oyster's free calculator to estimate what a global hire could really cost, so you can build a more accurate hiring budget and avoid surprises.