In partnership with

Ins7ghts Daily, by Yves Mulkers · Tuesday 17 November

Feeding it more is the expensive way to be wrong.

Three pieces from yesterday, one subject: how much you put in front of the model before it answers. Two of them are measurements and one is a product decision, and they do not point the same way.

Today's map. 17-23 points of accuracy: What a 4 KB file describing the data added, on three models. Without the 4 KB file 45.5%; With the 4 KB file 67.7%. Answers fell off past 2,500 tokens of retrieved context, and the bill kept running.

Advertisement

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

1. A 4 KB file beat the bigger context window

Airbyte published a comparison of the three ways companies let people ask questions of their own data: text to SQL, retrieval, and agents. Two measurements in it are worth keeping.

A controlled study on Natural Questions found that adding a 4 KB document describing what the data means raised accuracy by 17 to 23 points across three frontier models, from between 45.5% and 50.5% to between 67.7% and 68.7%.

The same write-up reports answer quality declining once the retrieved context passed roughly 2,500 tokens, and chunk overlap adding indexing cost with no measurable quality benefit.

Three dated pieces sit around it. On 12 November a data-management outlet called context the next great open-source battle. On 13 November AtScale argued that the fastest AI answer is usually the wrong one. Yesterday Tristan Handy of Fivetran and dbt talked about agents as the new data consumers and why dashboards stop being the product.

I read that 4 KB number as the cheapest data quality win on the table, and also as an indictment. Write down what a column means and the accuracy moves 20 points, and almost every team I have worked with kept that knowledge in one person's head and nowhere else. I would spend the next two weeks writing definitions before evaluating another retrieval stack.

Read the source →

 

Our research · Snowflake's real fight is not on its competitor list

Snowflake's real fight is not on its competitor list

What the field publishes about Snowflake shows who it really competes with, and that fight is not on its own competitor list.

Read the research →

Advertisement

Your buyers search in ChatGPT. Are you there?

Your buyers still search. They just do it in ChatGPT, Claude and Perplexity now. ChatFeatured shows where you rank in AI answers against your competitors, then writes and publishes the content that gets you named.

2. $368 of one team's bill came from checks that ran too late

A software engineering outfit published what its team actually pays for Claude Code, with the month broken out. The seat price is the dull part:

  • $709 went on extra rounds caused by failed builds and review comments.
  • $368 of that came from lint and spec-link checks that only ran after the agent had pushed its code.
  • Keeping context warm is billed on its own: a five-minute cache write costs 1.25 times the input rate, a one-hour cache twice.
  • Claude 4.7 and later tokenizers turn the same text into about 30% more tokens than the older ones did.

Yesterday InfoQ wrote up the other side of the ledger, Uber's legal agent cutting contract review time by 20%.

The model was not the expensive part of that month, the order of the work was. Checks that fire after the push are a pipeline problem, and in the quarterly review it will land on a line labelled AI, which is how a team ends up cutting the wrong thing. I would move every cheap check in front of the agent before I argued about a single rate.

Read the source →

 

Playbook · Stop your AI confusing Alphabet with alphabet

Stop your AI confusing Alphabet with alphabet

A quick name match told our system that Okta was a first name and Fidelity's holding company was a brain scan. The playbook shows how to stop AI mixing up companies that share a name.

Try the playbook →

3. Copilot will read the tabs you left open

Microsoft is rolling out a capability this month that lets the Microsoft 365 Copilot desktop app use the content of your open webpage tabs as context when it answers questions and writes summaries. The write-up tells administrators to go back over browser access, data governance and AI-tool interaction policies, because the assistant now reaches content that never passed through any of them.

Two things from yesterday sit next to that. Heimdal, which watches Windows fleets, has Copilot on 6 in 10 of the devices it monitors, after Microsoft resumed installing the app in the background on commercial Microsoft 365 desktops. The same day Arctera launched a connector that routes Copilot to enterprise data through its governance platform, so the controls stay on the path.

One vendor sold a governed path to the data and the platform widened the ungoverned one, on the same Monday. The tab open on my second screen at four in the afternoon is a customer record in a web app nobody ever classified, and that is the material this feature reads. I expect the next serious Copilot incident in a regulated firm to start in a browser tab rather than in a document library, and without logs nobody bought, it will not be provable either way.

Read the source →

Names to watch

Seen this week by Ins7ghts, our research platform: a knowledge graph of the data and AI news we read every day.

  • PxE Imaging · an Israeli startup whose Holocoder adds colour and depth capture to an ordinary image sensor
    Grounding. Metalenz said on 11 November that it had acquired the company to put colour and depth sensing into the perception platform behind its Polar ID sensor, and the optics industry body carried the announcement two days later.

Companies that later raised money had been in the news about twice as often in the quarter before. Worth a second look in three months.

Today's context on 7wData

Advertisement

What Creators Really Want. Straight From 300+ of Them.

This report shares insights from 300+ creators on what makes brand partnerships last.

  • Learn why 45% of creators want significant or complete creative control

  • Learn why 51% of creators need multiple collaborations before committing

  • Discover what creators say makes them commit long term

Build partnerships that renew instead of reset.

Also today

First MoveThe bug finder behind Cohere Cyber

Cohere's week reads as a security push: a unit called Cohere Cyber, a chief information security officer, a vice president for AI security. Step one company behind the announcement and you find what it actually bought. cubic, a code reviewer in London. AISLE, a bug finder in Prague that ran air-gapped with its results unpublished. Those two products sit in the gap between what a model writes and what gets merged, and the company that sells the writing now owns the gap.

Who signs off on your AI-written code, and do they also sell you the thing that wrote it?

My call: by 30 June 2027, at least one more large model vendor announces an acquisition of a code-review or application-security company.

Read the deep dive →

If you have measured what a written-down definition did to your answers, hit reply and tell me what it moved.

Yves

Tomorrow, Reality Check: a claim everyone repeats, held against what actually happened.

Know someone who should read this? Forward it with your link. Five people in, and the Signal vs Noise guide is yours.

Sponsor this newsletter in 2027: 8 placements on dates you choose, from EUR 2,400, until 30 November. Book: 7wdata.be/sponsor-2027

Ins7ghts Daily is published by 7wData.