@andrebergholz.bsky.social
#ChatGPT 's data analysis sandbox can't make external #API calls. #Claude 's container can, on by default on personal plans.
One setting decides whether you connect a service through an #MCP server or run its #CLI.
Five cases, and what I type in each: pub.towardsai.net/2785cd6a9efc
One setting decides whether you connect a service through an #MCP server or run its #CLI.
Five cases, and what I type in each: pub.towardsai.net/2785cd6a9efc
API for Volume, MCP for Chat, CLI for Terminals
Five common jobs, their corresponding commands, and the two checks that differentiate them
pub.towardsai.net
September 17, 2026 at 8:22 PM
Anthropic deleted 80% of Claude Code's system prompt. No measurable loss on coding evals.
Every instruction you write to patch a model's weakness has an expiry date. The tests that define "done" don't.
What survives when the model keeps changing:
aiadvances.org/4c7e4366a6a6
#AI #Claude #Software
Every instruction you write to patch a model's weakness has an expiry date. The tests that define "done" don't.
What survives when the model keeps changing:
aiadvances.org/4c7e4366a6a6
#AI #Claude #Software
Probably, Approximately, Shipped
Building software on a model that changes every six weeks breaks the one assumption the whole discipline was built on.
aiadvances.org
September 3, 2026 at 8:17 AM
Anthropic deleted 80% of Claude Code's system prompt. No measurable loss on coding evals.
Every instruction you write to patch a model's weakness has an expiry date. The tests that define "done" don't.
What survives when the model keeps changing:
aiadvances.org/4c7e4366a6a6
#AI #Claude #Software
Every instruction you write to patch a model's weakness has an expiry date. The tests that define "done" don't.
What survives when the model keeps changing:
aiadvances.org/4c7e4366a6a6
#AI #Claude #Software
I asked a chatbot a football rule I didn't know. Clean answer, totally confident and wrong. It had merged two opposite rulebooks into one smooth paragraph.
Fluent language isn't a sign of good reasoning. New piece:
medium.com/@bergholz/42...
#AI #LLMs
Fluent language isn't a sign of good reasoning. New piece:
medium.com/@bergholz/42...
#AI #LLMs
Too Smart to Be Right: Why Fluent Language Isn’t a Sign of Good Reasoning
A confident, well-phrased answer earns trust it hasn’t earned. From a machine, and from a person.
medium.com
August 17, 2026 at 9:33 AM
I asked a chatbot a football rule I didn't know. Clean answer, totally confident and wrong. It had merged two opposite rulebooks into one smooth paragraph.
Fluent language isn't a sign of good reasoning. New piece:
medium.com/@bergholz/42...
#AI #LLMs
Fluent language isn't a sign of good reasoning. New piece:
medium.com/@bergholz/42...
#AI #LLMs
#AI is barely 70 years old. In that short time it has buried its own best ideas at least twice and dug some of them back up.
I traced that #History through an unlikely lens: the eight times the #TuringAward has gone to #AI. New piece.
ai.gopubby.com/f79b12ab7699
I traced that #History through an unlikely lens: the eight times the #TuringAward has gone to #AI. New piece.
ai.gopubby.com/f79b12ab7699
The Turing Award’s AI Story
What computer science’s highest prize reveals about who shaped AI and what its silences leave unspoken
ai.gopubby.com
July 14, 2026 at 9:18 AM
#AI is barely 70 years old. In that short time it has buried its own best ideas at least twice and dug some of them back up.
I traced that #History through an unlikely lens: the eight times the #TuringAward has gone to #AI. New piece.
ai.gopubby.com/f79b12ab7699
I traced that #History through an unlikely lens: the eight times the #TuringAward has gone to #AI. New piece.
ai.gopubby.com/f79b12ab7699
A year ago I brought things to the #AI.
Now the AI is inside the things.
My workflow, one year later → levelup.gitconnected.com/5ebfba2934dc
#Productivity #Software #Development
Now the AI is inside the things.
My workflow, one year later → levelup.gitconnected.com/5ebfba2934dc
#Productivity #Software #Development
My AI Workflow, One Year Later
A year ago I said this would feel outdated in six months. I was right.
levelup.gitconnected.com
June 25, 2026 at 7:29 AM
A year ago I brought things to the #AI.
Now the AI is inside the things.
My workflow, one year later → levelup.gitconnected.com/5ebfba2934dc
#Productivity #Software #Development
Now the AI is inside the things.
My workflow, one year later → levelup.gitconnected.com/5ebfba2934dc
#Productivity #Software #Development
Benchmarks measure capability. Users experience behavior.
They're shaped by two completely different processes. And only one of them is fixed.
New piece on pre-training vs. post-training:
medium.com/generative-a...
#AI #LLMs #Training
They're shaped by two completely different processes. And only one of them is fixed.
New piece on pre-training vs. post-training:
medium.com/generative-a...
#AI #LLMs #Training
Pre-Training Gives LLMs Their Capability. Post-Training Gives Them Their Behavior.
The model you use has two biographies — and one of them you can help shape
medium.com
June 2, 2026 at 8:34 AM
Benchmarks measure capability. Users experience behavior.
They're shaped by two completely different processes. And only one of them is fixed.
New piece on pre-training vs. post-training:
medium.com/generative-a...
#AI #LLMs #Training
They're shaped by two completely different processes. And only one of them is fixed.
New piece on pre-training vs. post-training:
medium.com/generative-a...
#AI #LLMs #Training
I wrote about how we turned presentation creation into a reproducible Markdown-to-PDF pipeline: corporate design as context, Quarto/RevealJS for output, Docker for consistency.
Useful if your team still pays the hidden slide tax.
#AI #Productivity #Slides #Markdown
medium.com/ai-in-plain-...
Useful if your team still pays the hidden slide tax.
#AI #Productivity #Slides #Markdown
medium.com/ai-in-plain-...
Slides Are a Build Artifact Now
A reproducible pipeline from Markdown to boardroom-ready PDF
medium.com
May 26, 2026 at 6:12 AM
I wrote about how we turned presentation creation into a reproducible Markdown-to-PDF pipeline: corporate design as context, Quarto/RevealJS for output, Docker for consistency.
Useful if your team still pays the hidden slide tax.
#AI #Productivity #Slides #Markdown
medium.com/ai-in-plain-...
Useful if your team still pays the hidden slide tax.
#AI #Productivity #Slides #Markdown
medium.com/ai-in-plain-...
Yuval Harari made an observation I couldn't shake: more intelligence tends to produce more elaborate delusion, not less. The same cognitive architecture enables both. My new piece on what I think that means for AI:
generativeai.pub/0e1cc3a58d01
#AI #LLMs
generativeai.pub/0e1cc3a58d01
#AI #LLMs
No Intelligence Without Illusion
The same cognitive architecture that produces human intelligence also produces human delusion. We are now scaling both.
generativeai.pub
April 27, 2026 at 10:20 AM
Yuval Harari made an observation I couldn't shake: more intelligence tends to produce more elaborate delusion, not less. The same cognitive architecture enables both. My new piece on what I think that means for AI:
generativeai.pub/0e1cc3a58d01
#AI #LLMs
generativeai.pub/0e1cc3a58d01
#AI #LLMs
Just published my first article for Generative AI magazine: Language models learn by imitation. Reinforcement learning changes the signal to whether something actually worked. Why this matters, and where Karpathy's autoresearch fits in.
generativeai.pub/ac78019190d6
#AI #LLM #AutoResearch
generativeai.pub/ac78019190d6
#AI #LLM #AutoResearch
Beyond the Imitation Game: The Reinforcement Revolution
What changes when a machine stops learning from humans and starts learning from reality
generativeai.pub
April 13, 2026 at 7:30 AM
Just published my first article for Generative AI magazine: Language models learn by imitation. Reinforcement learning changes the signal to whether something actually worked. Why this matters, and where Karpathy's autoresearch fits in.
generativeai.pub/ac78019190d6
#AI #LLM #AutoResearch
generativeai.pub/ac78019190d6
#AI #LLM #AutoResearch
System prompts, commands, and skills look similar. They're not.
One is always-on context. One is a workflow you trigger. One is expertise the model applies on its own judgment.
#AI #Claude #Skills
pub.towardsai.net/af2fc985fa74
One is always-on context. One is a workflow you trigger. One is expertise the model applies on its own judgment.
#AI #Claude #Skills
pub.towardsai.net/af2fc985fa74
System Prompts, Commands & Skills: How AI Personalization Grew Up
Three ways to shape an AI assistant and why they’re not interchangeable
pub.towardsai.net
March 26, 2026 at 5:04 PM
System prompts, commands, and skills look similar. They're not.
One is always-on context. One is a workflow you trigger. One is expertise the model applies on its own judgment.
#AI #Claude #Skills
pub.towardsai.net/af2fc985fa74
One is always-on context. One is a workflow you trigger. One is expertise the model applies on its own judgment.
#AI #Claude #Skills
pub.towardsai.net/af2fc985fa74
Neural networks were declared dead twice. Both times, someone frustrated enough engineered a way through.
I traced 25+ architectures across 80 years. Six families of ideas, each born from a wall the previous generation couldn't get past.
medium.com/towards-arti...
#AI #History
I traced 25+ architectures across 80 years. Six families of ideas, each born from a wall the previous generation couldn't get past.
medium.com/towards-arti...
#AI #History
The Long Road to Overnight Success — Neural Nets from McCulloch-Pitts to Transformers
The forgotten ideas, dead ends, and stubborn breakthroughs behind today’s AI
medium.com
March 13, 2026 at 8:21 AM
Neural networks were declared dead twice. Both times, someone frustrated enough engineered a way through.
I traced 25+ architectures across 80 years. Six families of ideas, each born from a wall the previous generation couldn't get past.
medium.com/towards-arti...
#AI #History
I traced 25+ architectures across 80 years. Six families of ideas, each born from a wall the previous generation couldn't get past.
medium.com/towards-arti...
#AI #History
"Just Google it" sounds dated now.
I ask AI instead. The real shift is the follow-up questions. That's where understanding happens.
New post on how iteration changed the way I learn:
ai.gopubby.com/7f4c45d86d57
I ask AI instead. The real shift is the follow-up questions. That's where understanding happens.
New post on how iteration changed the way I learn:
ai.gopubby.com/7f4c45d86d57
Deep Research vs. Broad Research: How AI turned Information Retrieval into Personalized Learning
On iteration, depth, and learning at your own pace
ai.gopubby.com
February 19, 2026 at 9:08 AM
"Just Google it" sounds dated now.
I ask AI instead. The real shift is the follow-up questions. That's where understanding happens.
New post on how iteration changed the way I learn:
ai.gopubby.com/7f4c45d86d57
I ask AI instead. The real shift is the follow-up questions. That's where understanding happens.
New post on how iteration changed the way I learn:
ai.gopubby.com/7f4c45d86d57
Your AI has been building a psychological profile of you.I asked mine to show it. Uncomfortably accurate. Here's the prompt + what it means 👇
ai.plainenglish.io/7863e640ea61
ai.plainenglish.io/7863e640ea61
Your AI Has Assembled a Profile of You: Here’s What It Knows
I asked AI to profile me, what came back was uncomfortably accurate
ai.plainenglish.io
February 5, 2026 at 8:10 AM
Your AI has been building a psychological profile of you.I asked mine to show it. Uncomfortably accurate. Here's the prompt + what it means 👇
ai.plainenglish.io/7863e640ea61
ai.plainenglish.io/7863e640ea61
New: State of the AI: January 2026
- GPT‑5.2’s reasoning jump
- Agentic coding pilots everywhere
- Token prices falling fast
- EU GPAI rules now real procurement constraints
A hands-on briefing for 2026 planning, published in "Towards AI" magazine.
medium.com/towards-arti...
- GPT‑5.2’s reasoning jump
- Agentic coding pilots everywhere
- Token prices falling fast
- EU GPAI rules now real procurement constraints
A hands-on briefing for 2026 planning, published in "Towards AI" magazine.
medium.com/towards-arti...
State of the AI: January 2026 Report
From pilot to production: Tracking the shift to agentic systems and the commoditization of reasoning
medium.com
January 13, 2026 at 11:05 AM
New: State of the AI: January 2026
- GPT‑5.2’s reasoning jump
- Agentic coding pilots everywhere
- Token prices falling fast
- EU GPAI rules now real procurement constraints
A hands-on briefing for 2026 planning, published in "Towards AI" magazine.
medium.com/towards-arti...
- GPT‑5.2’s reasoning jump
- Agentic coding pilots everywhere
- Token prices falling fast
- EU GPAI rules now real procurement constraints
A hands-on briefing for 2026 planning, published in "Towards AI" magazine.
medium.com/towards-arti...
The idea that started AI: How 4 researchers in 1956 bet thinking was computable and launched 70 years of progress. Dartmouth Conference story now live on Medium.
➡️ medium.com/ai-in-plain-...
#AIHistory #McCarthy #Minsky #ArtificialIntelligence
➡️ medium.com/ai-in-plain-...
#AIHistory #McCarthy #Minsky #ArtificialIntelligence
The Idea That Started AI and the Researchers Who Built It
In the summer of 1956, a small group of researchers convinced themselves that thinking was computable and that eight weeks could prove it
medium.com
December 17, 2025 at 9:58 AM
The idea that started AI: How 4 researchers in 1956 bet thinking was computable and launched 70 years of progress. Dartmouth Conference story now live on Medium.
➡️ medium.com/ai-in-plain-...
#AIHistory #McCarthy #Minsky #ArtificialIntelligence
➡️ medium.com/ai-in-plain-...
#AIHistory #McCarthy #Minsky #ArtificialIntelligence
Excited to announce my first article for Towards AI magazine:
I built an "Omnibar AI" that connects my command line directly to #OpenAI using #n8n. For zero-friction intelligence at your fingertips.
Here is the full architectural guide:
medium.com/towards-arti...
#AI #Automation #Productivity
I built an "Omnibar AI" that connects my command line directly to #OpenAI using #n8n. For zero-friction intelligence at your fingertips.
Here is the full architectural guide:
medium.com/towards-arti...
#AI #Automation #Productivity
The Omnibar Pattern: A Simple Architecture for Real-World AI
Don’t lock your AI in a browser tab. Build a utility that lives at your fingertips.
medium.com
December 8, 2025 at 1:39 PM
Excited to announce my first article for Towards AI magazine:
I built an "Omnibar AI" that connects my command line directly to #OpenAI using #n8n. For zero-friction intelligence at your fingertips.
Here is the full architectural guide:
medium.com/towards-arti...
#AI #Automation #Productivity
I built an "Omnibar AI" that connects my command line directly to #OpenAI using #n8n. For zero-friction intelligence at your fingertips.
Here is the full architectural guide:
medium.com/towards-arti...
#AI #Automation #Productivity
AI-first = adding intelligence
AI-native = rebuilding the business around it
One optimizes, the other rewrites the rules. Which one are you really?
#AI #ArtificialIntelligence #AINative #AIFirst #Strategy
pub.aimind.so/4f9c437921f8
AI-native = rebuilding the business around it
One optimizes, the other rewrites the rules. Which one are you really?
#AI #ArtificialIntelligence #AINative #AIFirst #Strategy
pub.aimind.so/4f9c437921f8
AI-First vs. AI-Native: The Distinction Your Company’s Future Depends On
Retrofit with AI or rebuild around it: which are you, and what should you do next?
pub.aimind.so
November 19, 2025 at 4:37 PM
AI-first = adding intelligence
AI-native = rebuilding the business around it
One optimizes, the other rewrites the rules. Which one are you really?
#AI #ArtificialIntelligence #AINative #AIFirst #Strategy
pub.aimind.so/4f9c437921f8
AI-native = rebuilding the business around it
One optimizes, the other rewrites the rules. Which one are you really?
#AI #ArtificialIntelligence #AINative #AIFirst #Strategy
pub.aimind.so/4f9c437921f8
Ever spent hours tweaking prompts with no improvement? My first piece for @AIAdvances shows you proven frameworks (COSTAR, TCREI) and a plug-and-play generator template. Stop guessing, start engineering.
ai.gopubby.com/cefdc03085b3
#AI #PromptEngineering
ai.gopubby.com/cefdc03085b3
#AI #PromptEngineering
Crafting Prompts That Click: A Framework-Driven Guide (Plus Your Ready-to-Use Prompt Generator…
Master proven frameworks like COSTAR and TCREI, plus get a plug-and-play prompt generator template for modern AI systems
ai.gopubby.com
November 5, 2025 at 3:25 PM
Ever spent hours tweaking prompts with no improvement? My first piece for @AIAdvances shows you proven frameworks (COSTAR, TCREI) and a plug-and-play generator template. Stop guessing, start engineering.
ai.gopubby.com/cefdc03085b3
#AI #PromptEngineering
ai.gopubby.com/cefdc03085b3
#AI #PromptEngineering
Explain Like I'm Human: Asking a neural network to break down billions of parameters is like a 5-year-old demanding you explain quantum physics.
My latest for @inPlainEngHQ :
ai.plainenglish.io/2a80ede090bb
#AI #ExplainableAI #MachineLearning
My latest for @inPlainEngHQ :
ai.plainenglish.io/2a80ede090bb
#AI #ExplainableAI #MachineLearning
Explain Like I’m Human: The Reality Check We Need About AI Explainability
The cognitive mismatch between human expectations and AI reality and why understanding machines may be like a five-year-old trying to…
ai.plainenglish.io
October 28, 2025 at 1:22 PM
Explain Like I'm Human: Asking a neural network to break down billions of parameters is like a 5-year-old demanding you explain quantum physics.
My latest for @inPlainEngHQ :
ai.plainenglish.io/2a80ede090bb
#AI #ExplainableAI #MachineLearning
My latest for @inPlainEngHQ :
ai.plainenglish.io/2a80ede090bb
#AI #ExplainableAI #MachineLearning
Just published my first article in @aimindlabs! 🚀
How did AI go from predicting "butter" after "bread and" to solving Math Olympiad problems? Reasoning models learned to automatically do what skilled humans were doing manually.
medium.com/ai-mind-labs...
#AI #PromptEngineering
How did AI go from predicting "butter" after "bread and" to solving Math Olympiad problems? Reasoning models learned to automatically do what skilled humans were doing manually.
medium.com/ai-mind-labs...
#AI #PromptEngineering
Reasoning Models Are Just Internalized Prompt Engineering: How AI Learned to Question Itself
How four breakthroughs taught AI systems to automatically do what skilled prompt engineers were already doing manually.
medium.com
October 15, 2025 at 8:16 AM
Just published my first article in @aimindlabs! 🚀
How did AI go from predicting "butter" after "bread and" to solving Math Olympiad problems? Reasoning models learned to automatically do what skilled humans were doing manually.
medium.com/ai-mind-labs...
#AI #PromptEngineering
How did AI go from predicting "butter" after "bread and" to solving Math Olympiad problems? Reasoning models learned to automatically do what skilled humans were doing manually.
medium.com/ai-mind-labs...
#AI #PromptEngineering
Made a simple guide through everything I've written about #AI so far. Four pillars: systems, intelligence, work, and history.
Find your starting point: medium.com/@bergholz/9a...
Find your starting point: medium.com/@bergholz/9a...
My AI Writing Compass
Everything I’ve Written About AI (So Far)
medium.com
September 24, 2025 at 1:37 PM
Made a simple guide through everything I've written about #AI so far. Four pillars: systems, intelligence, work, and history.
Find your starting point: medium.com/@bergholz/9a...
Find your starting point: medium.com/@bergholz/9a...
Which AI model is ‘good enough’ for the job? This guide breaks down model selection, cost, and creative fit—plus a visual workflow cheatsheet for busy teams:
medium.com/@bergholz/b6...
medium.com/@bergholz/b6...
Model Choice in the GPT-5 Era: Spend Smart, Ship Fast, Defer the Fancy
GPT-5 is the default; model choice still matters for cost, speed, and special features.
medium.com
September 16, 2025 at 12:58 PM
Which AI model is ‘good enough’ for the job? This guide breaks down model selection, cost, and creative fit—plus a visual workflow cheatsheet for busy teams:
medium.com/@bergholz/b6...
medium.com/@bergholz/b6...
My ChatGPT conversations start with one thing: a well-crafted system prompt.
Here’s the one I use, why it works, and how I keep it sharp.
ai.plainenglish.io/7e926b8396d0
#ChatGPT #PromptEngineering #AI
Here’s the one I use, why it works, and how I keep it sharp.
ai.plainenglish.io/7e926b8396d0
#ChatGPT #PromptEngineering #AI
My ChatGPT System Prompt (with GPT-5 Update)
Copy and tweak as you like.
ai.plainenglish.io
August 14, 2025 at 1:19 PM
My ChatGPT conversations start with one thing: a well-crafted system prompt.
Here’s the one I use, why it works, and how I keep it sharp.
ai.plainenglish.io/7e926b8396d0
#ChatGPT #PromptEngineering #AI
Here’s the one I use, why it works, and how I keep it sharp.
ai.plainenglish.io/7e926b8396d0
#ChatGPT #PromptEngineering #AI
AlphaGo’s historic victory over Lee Sedol was more than another win for AI. It was a first glimpse into the future of machine creativity. My latest Medium article explores what AlphaGo really changed. Read here:
ai.plainenglish.io/3dfbc0bc0efb
#AI #AlphaGo #DeepMind #GoGame
ai.plainenglish.io/3dfbc0bc0efb
#AI #AlphaGo #DeepMind #GoGame
Beyond Calculation: AlphaGo and the Birth of Machine Creativity
How one move on a 19×19 board rewrote the rules of artificial intelligence
ai.plainenglish.io
August 4, 2025 at 3:18 PM
AlphaGo’s historic victory over Lee Sedol was more than another win for AI. It was a first glimpse into the future of machine creativity. My latest Medium article explores what AlphaGo really changed. Read here:
ai.plainenglish.io/3dfbc0bc0efb
#AI #AlphaGo #DeepMind #GoGame
ai.plainenglish.io/3dfbc0bc0efb
#AI #AlphaGo #DeepMind #GoGame
🚀 New on Medium and in AI in Plain English: “Truth, Trust, and Hallucinations” @inPlainEngHQ
Explore how modern AI balances accuracy and trustworthiness, plus strategies for spotting and reducing hallucinations.
Read now → medium.com/@bergholz/17...
#AI #LLMs
Explore how modern AI balances accuracy and trustworthiness, plus strategies for spotting and reducing hallucinations.
Read now → medium.com/@bergholz/17...
#AI #LLMs
Truth, Trust, and Hallucinations
How to navigate the fine line between useful creativity and confident nonsense.
medium.com
July 23, 2025 at 2:43 PM
🚀 New on Medium and in AI in Plain English: “Truth, Trust, and Hallucinations” @inPlainEngHQ
Explore how modern AI balances accuracy and trustworthiness, plus strategies for spotting and reducing hallucinations.
Read now → medium.com/@bergholz/17...
#AI #LLMs
Explore how modern AI balances accuracy and trustworthiness, plus strategies for spotting and reducing hallucinations.
Read now → medium.com/@bergholz/17...
#AI #LLMs