Lucas van Walstijn

[Lucas van Walstijn @lvWal]

@randal_olson From afar it looks good, but pixel level image integrity is extremely low, also crazy many image errors! Heads being mounted to the back of a torso.. legs not making sense.. No artist would ever create such slop

[Lucas van Walstijn @lvWal]

@gvanrossum Saying that agents are just LLMs plus tool calls is a bit like describing a computer as electricity flowing through circuits to process and store information. It’s technically accurate, but it completely undersells what makes it powerful.

[Lucas van Walstijn @lvWal]

@jlowin Hey @jlowin this is **very** cool. I am trying to get this working in Langchains deepagents framework. What's the magic you are using for the html to actually render as such in the frontend? Do you have any documentation of how you set up this demo?

[Lucas van Walstijn @lvWal]

@trq212 This is awesome, thanks for sharing. Do you have perhaps any pointers for techniques to show charts (or other artifacts) interleaved with its text output to users?

[Lucas van Walstijn @lvWal]

@ph_singer @sparbuchfeinde Fake news..

1️⃣ Regime für ausländische Rentner (7 %-Regel)

- 7 % Flat Tax
- nur auf ausländische Renteneinkünfte
- Nur in kleinen Gemeinden in Süditalien
- Dividenden sind ausdrücklich NICHT begünstigt

[Lucas van Walstijn @lvWal]

@rohanpaul_ai @ylecun His 180 degrees is completed 😅
https://t.co/YjN0NAYYSN

[Lucas van Walstijn @lvWal]

@ElstnerMartin @ginandstocks I know of no other population that loves to bash itself as much as the Germans.

[Lucas van Walstijn @lvWal]

@HamelHusain Cool! I hope she likes it! Here's my honest feedback on the "easy division" task.. pedagogically, I don’t think it works very well. The UI feels unintuitive, and in my opinion it’s not a great approach for helping learners really understand division. ¯\_(ツ)_/¯

[Lucas van Walstijn @lvWal]

@XandervdWulp @NOSsport kunnen jullie ajb de presentatoren informeren dat winterberg gewoon winterberg heet en niet WIENTERberg. Vooral bij mevr de Zeeuw en mevr de Graaf blijkt verwarring te bestaan wat betreft de correcte uitspraak..

[Lucas van Walstijn @lvWal]

@abhi1thakur Sorry to hear, I have send you a DM

[Lucas van Walstijn @lvWal]

@abhi1thakur What are you using NotebookLM for?

[Lucas van Walstijn @lvWal]

@jeremyphoward @Teknium Although your statement is true, I don't think this is the best analogy. Deep Blue was created to do chess, not to do text extraction. There is more nuance to OPs message. The analogy to Deep Blue: why can it do great openings, but sucks at the end game?

[Lucas van Walstijn @lvWal]

@Teknium Could it be effective to basically fine-tune on exactly that signal? In prod: whenever a user tells wtf the model did wrong, let the model also feedback on it's own output, and update weights on difference between model and user feedback?

[Lucas van Walstijn @lvWal]

@simonw Could it be that sharing the (bad) results of these models over time, pollute the training data of newer models? Especially since there is practically no other training data to be found on the internet. (not meant as a critique to your benchmark..)

[Lucas van Walstijn @lvWal]

@teknium For some reason it seems there is a new type of ghost text. It's added automatically in all lines, is grey and has a "double click to add" tooltip. Pressing ESC doesn't make them disappear.. pretty annoying. Seems to only happen when using partial

[Lucas van Walstijn @lvWal]

@fchollet As well as @jeremyphoward teaching style with #FastAI

[Lucas van Walstijn @lvWal]

@fchollet Fully agree. Reminded me of the famous Feynman quote: "What I cannot build. I do not understand"

[Lucas van Walstijn @lvWal]

@abhi1thakur @iScienceLuvr thnx!

[Lucas van Walstijn @lvWal]

@abhi1thakur @iScienceLuvr Would you care to answer the question? I am also interested to learn more about your observation :)

[Clémentine Fourrier 🍊 is off till Dec 2026 (🪂) @clefourrier]

If you can't shell out 2K$ 😱 to learn about LLM evaluations, take a look at our free/open resources:
1. LLM guidebook: From theory to troubleshooting
https://t.co/sLzdTx8VW9
2. YourBench: generate evals from your custom data, with no code and no effort
https://t.co/Vg5Df65lIh