Do LLMs fall for the gambler’s fallacy? “Most of these models do not. For long streaks they do roughly the opposite.”
@mangal_utkarsh reads small models’ next-token probabilities to see what they really expect after a run of heads. towardsdatascience.com/after…
844
It’s your tenth AI approval request of the afternoon, and you click approve. “Technically there was a ‘human in the loop’, but was there really?”
Corné Potgieter draws on Kahneman’s System 1 and System 2 thinking to show why oversight fails, and how to design loops that work. towardsdatascience.com/why-h…
How do you avoid becoming “a human whose only job is to pass requests to an AI and its answers back”?
Drawing on manufacturing and logistics projects, @Samir_Saci_ shares the skills that keep analytics professionals valuable as AI agents take on more of the work. towardsdatascience.com/how-t…
1
1,083
Is a green test suite worth anything when the same AI wrote the code and the tests? Gal Arav built qikly to keep them apart. The coding agent never sees the criteria the tests are written from.
towardsdatascience.com/qikly…
1
1,138
Running several coding agents at once? You’ve probably lost time hopping between sessions to see which one is waiting on you.
@EivindKjos shares a simple fix: a central “herder” agent that collects every pending question and surfaces them by priority. towardsdatascience.com/how-t…
2
3
1,522
“At that point, the challenge is no longer prompting. It is knowledge governance.”
@erikapatg shows how to package coding-agent skills, rules, and hooks as versioned plugins so teams can share them across repos without drift. towardsdatascience.com/how-t…
Your AI agent doesn’t need you staring at a chat window while it works. Benjamin Etienne walks us through webhooks, persistent sessions, and human-in-the-loop checkpoints for long-running agents in Google ADK. towardsdatascience.com/googl…
3
6
1,763
“The trouble is that they add up to a second application hiding behind a protocol endpoint.”
In the first part of a series on running MCP systems, Priyansh Bhardwaj explains why servers should expose capabilities and leave planning, state, and permissions to the host. towardsdatascience.com/how-t…
Want an AI assistant that answers questions about your own library, cites the exact page, and runs on your laptop? @WaleedEsmail13 Waleed Esmail shares how he built one, how he tested it, and what he learned about small local models along the way. towardsdatascience.com/local…
A test agent and a coding agent, kept apart so a green suite means something
By Gal Arav
towardsdatascience.com/qikly…
1
7
1,603
What if your knowledge graph remembered past decisions, user preferences, and how incidents were resolved, not just static facts? Partha Sarkar walks us through six agentic GraphRAG patterns for turning a passive graph into a living source of enterprise context. towardsdatascience.com/a-pra…
“Just because we can model something as a Markov chain doesn’t mean we necessarily should.”
Ananya Bhattacharyya’s beginner-friendly guide shows how Markov chains work in NumPy, and how to tell when the “memoryless” assumption breaks down. towardsdatascience.com/how-t…
Running AI agents in production? If your monitoring starts with error rates, you may be watching the last thing to change.
@BarrMoses_MC explains why quality, cost, and latency checks catch problems sooner, and where most teams have blind spots. towardsdatascience.com/data-…
Counting parameters doesn’t tell you how much a quantum model can really learn.
Adrian Panasiewicz and Jacob L. Cybulski turn the theory of effective dimension into working code, with practical tips for keeping the math numerically stable. towardsdatascience.com/compu…
Towards Data Science retweeted
My @TDataScience article: a local study assistant on one 16 GB GPU, graded blind.
Small local models: 12–16/20 on textbook questions, 0–2/15 on a 2026 paper. Same passages.
I hope you find it helpful, and I'd appreciate contributions!
towardsdatascience.com/local…
From protecting API keys to keeping track of context usage, small safeguards can make AI coding workflows more reliable. Chien Vu Minh explains how to build and test four custom Claude Code mods to address common frustrations.
towardsdatascience.com/claud…
7
1,911
What happens when a RAG system understands the meaning of a query but misses the exact identifier it needs? Spyros Georgopoulos looks at how a 1994 algorithm could help address this retrieval gap.
towardsdatascience.com/your-…
1
4
1,777
AI agents need more than prompts to work effectively; they need well-designed harnesses. Sam Black breaks down three foundational concepts to help engineers and data scientists prepare for this shift.
towardsdatascience.com/what-…
Towards Data Science retweeted
JEV can spot tiny differences in similar images. Check out the example below and my @TDataScience article at track.pstmrk.it/3s/towardsda…
Hint: look for a faint horizontal scratch in the middle of the door on the right
Towards Data Science retweeted
Your AI agent writes the code and the tests, but who checks the tests?
My 10-min walkthrough on @TDataScience: qikly keeps the agents apart and I explain how.
▶️ youtube.com/watch?v=paNIbTQP…
Try it on your own code: pip install qikly github.com/gal-a/qikly




