Sponsored by
Unify Your Teams and Tech Stack With HubSpot
Connect your customer data, teams, and tools without the hassle of complex integrations or lengthy setups. One easy platform gives marketing, sales, and service teams a unified customer view and the tools to turn it into growth.
Why HubSpot and what's new
• Generate leads and automate marketing with Marketing Hub
• Build your pipeline and close more deals with Sales Hub
• Scale customer support and drive retention with Service Hub
• Keep customer data clean, connected, and actionable with one, unified platform
Join 306,000+ in over 135 countries using HubSpot to grow their businesses.
See what a more connected approach to growth can do for you and your team. Get setup quickly and start checking off your hardest tasks.
Sponsor:
iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #157 WEDNESDAY · 14 OCTOBER 2026
THE HOOK
On 6 October, OpenAI posted more than 700 maths manuscripts to GitHub. They came from about 4,000 problems it set a model it hasn’t released. A spokesperson told Scientific American that nearly all the results came from one prompt, handed to one agent, rather than instructions written for each problem. The prompt wasn’t published. Then Scott Aaronson wrote that ‘no human has understood just about any of these proofs yet’.
OUR ANGLE
Checking is the new bottleneck
OpenAI’s repository shows the split. About 42% of the top-line results have a Lean formalisation, where software verifies each step mechanically. That settles the formal statement, not necessarily everything the paper claims. The rest, the README admits, ‘could have issues’, and they wait for mathematicians who’ve only just started reading.
Software has the same queue. In Undo’s survey of 300 senior engineering leaders at large US and UK firms building mission-critical software, 81% linked AI coding tools to a production incident in the past six months, and 79% said releases are no faster than before AI.
My read: whether AI output is worth using now depends heavily on what it costs to check. Where a machine can check it, the gains stack up. Where only a person can, the work queues in front of them. So before your next prompt, ask how you’ll know it worked, and how quickly.
Prediction: by 30 June 2027, OpenAI’s README will report at least 75% of top-line results formalised in Lean, up from about 42% on 14 October. OpenAI says more formalisations are coming; the hardest proofs will lag.
Go deeper: Build the check before you ask follows one recurring task through a failed check, a fix and a re-test.
ALSO THIS WEEK
Claude’s smallest model got much cheaper
Anthropic released Claude Haiku 5.5 on 7 October. For requests under 100,000 tokens, list prices fall 90%, to $0.10 per million tokens in and $0.50 out. Anthropic’s headline 75% is its estimate of the average saving once longer requests and a new tokenizer are counted. Checking the output isn’t on the price list.
Microsoft’s CEO: assume the model is compromised
‘We must assume a model is compromised and contain it from the start,’ Satya Nadella posted on X on Saturday. He wants an emergency brake for pausing a model mid-task, and tamper-proof records of what agents do. Note who’s saying it: a company that sells these models.
Send your AI twin to the interview?
Eightfold AI’s chief executive, Ashutosh Garg, has launched TwinCafe. Job seekers upload a CV, LinkedIn profile and work samples, and recruiters interview the resulting AI twin before meeting the person. Live interviews, Garg says, can then focus on ‘the human aspect versus the fact-check aspect’. Which leaves an obvious question: who fact-checks the twin?
PROMPT OF THE WEEK
Write the test before the task
Use it before you hand AI a job you’ll repeat: coding expenses, summarising contracts, cleaning a list. It drafts the checks; you choose which to keep. Do the job afterwards, in a fresh chat.
Before any work is done, help me define how I’ll know the result is right. Don’t do the task itself.
TASK: [what I’ll ask an AI to do, e.g. code 400 expense lines to our ledger accounts]
INPUTS: [what I’ll give it]
USED FOR: [the decision or document it feeds]
1. List checks a machine or formula could run on the output: totals that must reconcile, counts, allowed values, required fields, formats.
2. List checks only a person can make, and which items to sample for them.
3. Suggest three known-answer cases I can take from past work and mix in unlabelled.
4. Describe two realistic mistakes I could plant in a copy of the output to test these checks.
5. Say what result should stop the work, e.g. fewer than 38 of 40 known answers correct.
Why it works: checks written after you’ve seen an answer drift towards it, because the answer quietly becomes the yardstick. Written first, they measure the output against what you needed.
TOOL OF THE WEEK
ChatGPT’s interactive answers, used as a test bench
A calculator you can poke is easier to check than a number you can’t.
Since 7 October, GPT-6 can answer in ChatGPT’s Chat tab with working parts, which OpenAI calls Intelligent UI: calculators, editable charts, forms and maps. It’s on every plan, though not in the Work tab or at the Pro reasoning setting.
Use it as a test bench. Ask for a calculator, then feed it a case you’ve already worked out by hand, such as last quarter’s figures. A mismatch exposes a wrong assumption in seconds. Then push one input to an extreme and see whether the answer still makes sense.
Verdict: worth five minutes on budgets, pricing and forecasts. A polished widget looks more trustworthy than it is, so ask for the formula in words, and trust it only once it matches your hand-worked case.
TIP OF THE WEEK
Plant a mistake to test your check
Every check has a blind spot: it only catches what it was built to catch. So test the check.
Take a copy of an AI output you’d normally approve and plant two realistic errors, say a duplicated line and a figure out by a factor of ten. Run your usual review. If a planted error gets through, real ones probably are too.
It works because you know where the errors are, so you’re measuring the review, not the AI. The limit: catching both shows the check works on those two mistakes, nothing more.
Pro move: have a colleague plant them, and don’t say how many.
YOUR MOVE
Build one check this week
Pick one AI task you’ll repeat this month. Run the prompt, keep two checks you’d actually use, and plant one error to see whether they catch it.
REPLY WITH ONE LINE
What did you check, and did the planted error get through? One sentence is plenty, for example: ‘Expense coding: our total check missed a planted duplicate.’ Leave out names and confidential details. I read every reply; we won’t publish a tally.
Know someone who signs off AI-assisted work? Forward them this.
Disclosure: the original draft used Claude. Anthropic makes Haiku 5.5, which appears above.
P.S. The companion shows what a passing score does and doesn’t prove, using one quarter’s invoices.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
iPROMPT / 157 /