Language Models

What exactly is a "weight" in an AI model?

We explain gradient descent, hyperparameters and overfitting, but almost never the thing that actually gets adjusted: the weight. It's a number, and a model is...

23 July 2026 7 min

Editor's picks

Language Models

'Open' in AI doesn't mean what you think: how to check it yourself

'Open weights' is not 'open source,' and the difference has real consequences: from user thresholds to territorial exclusions buried in the fine print. Using Llama 4's actual license and the counterexample of OLMo 2, this piece teaches three questions to check for yourself, document in hand, whether an AI model billed as 'open' really is.

8 min 23 Jul. 2026

Language Models

What it actually takes to train your own model: the math nobody shows you

"How much GPU do you need" is the wrong question. Compute, data and knowledge don't add up — they're three gates in series, and the one almost nobody clears isn't the hardware one. With today's real pricing, a real open model's corpus, and the real logbook from a 175-billion-parameter training run, this piece does the full math — and tells you which gate you're actually standing at.

7 min 25 Jul. 2026

Practical applications

How to verify a chatbot claim before you use it

A practical method for turning a persuasive AI answer into a checkable claim: isolate the fact, find the primary source, and decide whether it belongs in a report, email, or decision.

9 min 26 Jul. 2026

Current Affairs

This Isn't a Launch: It's Anthropic's Quiet Lesson in Talking to Any AI

Headlines are circulating about a supposed launch of Claude Code's prompt library. Verified with the Wayback Machine: it has existed since May, not this week. There's no news to report, but there are four principles — pulled from Anthropic's own documentation — for judging any prompt, for any AI, that will still hold once this tool is gone.

8 min 26 Jul. 2026

Latest articles

Current Affairs

We Asked Our Own March Piece for Its Sources. Here's What We Found

In March we wrote a draft — never published, sitting in the queue for four and a half months — citing six percentages with no link at all: four attributed to 'a Harvard Business Review study' and two to a Harness survey. We checked them: the four from HBR aren't in the real study, the two from Harness do exist, and the 'unnamed engineer' had a name all along. The method left over for any unsourced number.

8 min 26 Jul. 2026

Education

An AI Detector Said a 30-Year-Old Essay Was Written by AI. Here's How to Actually Read One of These Accusations

A Yale student proved GPTZero gave a '100% probability of AI' verdict on texts published decades before any language model existed. A Stanford study measured the problem precisely: 61% false positives on real student writing. The four questions worth asking before accepting any accusation like this.

7 min 26 Jul. 2026

AI Fundamentals

When several AI models agree, they may still be wrong

A new preprint warns that self-consistency and agreement between LLMs are weak, context-dependent confidence signals.

2 min 26 Jul. 2026

Current Affairs

Your Loyalty Card Isn't Just a Discount Anymore — It's How an AI Knows It's You

Sephora let customers link their Beauty Insider account inside ChatGPT in March 2026; Ulta did the same with Google Gemini in April, covering 46 million loyalty members. It isn't unique to beauty retail: it's the same mechanism feeding any AI shopping assistant. What you can demand about that data today, and what's still unresolved in every state.

8 min 26 Jul. 2026

Use Cases

Estonia tests AI for finding errors in draft laws

A beta prototype analyses legislative texts to flag broken references, contradictions and impossible dates before a bill is approved.

4 min 26 Jul. 2026

AI Fundamentals

How to read an AI paper without being a researcher

Every summary of OPT-175B's training said the same thing: 992 GPUs, two months. None mentioned the 35 manual restarts at three in the morning. They were sitting in the actual document, under a plainly-named section almost nobody opens. This piece teaches which parts of a paper to read, in what order, and walks the exact route — minutes, not a research career — through the real case that changed the argument of my previous piece.

7 min 26 Jul. 2026

Medicine

A proposal for assessing AI safety in mental health

A new preprint argues that language models used for emotional support should be assessed through explicit values, training and ongoing oversight.

4 min 25 Jul. 2026

Current Affairs

Swisscanto says it is too early to call AI a bubble

Gerhard Wagner of Swisscanto separates AI’s structural potential from specific overvaluations and from the challenge of turning investment into profits.

4 min 25 Jul. 2026

Current Affairs

Memora proposes a different kind of memory for AI agents

Microsoft Research separates what an agent stores from how it retrieves it, aiming to handle long histories without loading the entire context.

3 min 25 Jul. 2026

What exactly is a "weight" in an AI model?

Editor's picks

'Open' in AI doesn't mean what you think: how to check it yourself

What it actually takes to train your own model: the math nobody shows you

How to verify a chatbot claim before you use it

This Isn't a Launch: It's Anthropic's Quiet Lesson in Talking to Any AI

Latest articles

We Asked Our Own March Piece for Its Sources. Here's What We Found

An AI Detector Said a 30-Year-Old Essay Was Written by AI. Here's How to Actually Read One of These Accusations

When several AI models agree, they may still be wrong

Your Loyalty Card Isn't Just a Discount Anymore — It's How an AI Knows It's You

Estonia tests AI for finding errors in draft laws

How to read an AI paper without being a researcher

A proposal for assessing AI safety in mental health

Swisscanto says it is too early to call AI a bubble

Memora proposes a different kind of memory for AI agents

Explore categories

Use Cases

AI Fundamentals

Artificial Intelligence Glossary

Current Affairs

Recommended Books

Practical applications

Regulatory Framework

Grok