Vojta Kopal

@sichy @MichalHudecek Jojo cele rodicovstvi outsourcovat na skolu, at maji rodice klid na praci.

Cist a psat se uci ve skole. Je jim teprve 6. Analyzu bych nechal na gympl. Ucitele at si to nezjedodusuji s nauci deti pekne cist, psat a pocitat od uplneho zacatku. A psat tim myslim krasopis a opet nezjedodusovst na ten nesmysl s komeniascriptem. Helveticy si uziji vic nez dost pak v appkach.

Poprvé v historii ČR máme k dispozici data o tom, co děti doopravdy umí, když přichází do první třídy. Učitelé z více než 1100 ZŠ provedli v aplikaci Včelka diagnostiku raných matematických dovedností s 25 000 žáky, tedy přibližně 1/4 všech nastupujících prvňáčků. 👇 Výsledky

Learn programming by playing games: Blockchain → cryptozombies.​io SQL → mystery.knightlab.​com JavaScript → warriorjs.​com Go → codingame.​com Python → codedex.​io CSS → cssbattle.​dev Java → tynker.​com Git → ohmygit.​org

Is that how living in the EU will feel in the Age of AI? 🤔 @AIatMeta's new multimodal Llama, @Apple Intelligence, or @OpenAI's Advanced Voice Mode of ChatGPT are all currently restricted in the EU. There's a growing concern that this could lead to "two-speed AI" - more advanced AI for the rest of the world and less capable AI for the EU. While regulations aim to protect, they can also hinder innovation—we need to quickly find a balance to ensure progress isn't stalled. 🇪🇺

@GergelyOrosz At @MewsRnD we did AI hackathon aiming to discover opportunities across our product portfolio. There were some great ideas and prototypes. We've also seen in the past hackathons examples when we tried to push the winning solution into production, that it was shortly deprecated.

Have you run a successful hackathon or observed a well-organized one? I'm writing up stories that can be helpful for teams and startups organizing a hackathon (eg hacking around AI APIs is relevant!) If you have input on this topic, happy to hear - both in DMs or replies.

@adent U rodičů-učitelů a konkrétních škol vidím, kde jsou ty problémy a kolik úvazků nebude k dispozici a jaký dopad to bude mít na výuku. Buď si to ředitelé neumí spočítat a Fiala neumí vysvětlit, a nebo fakt je Vláda tak mimo, že jim stačí, když to vychází na průměrnou(!) školu.

Přesně tak. O učitele a jejich platy tam vůbec nešlo. Šlo, jako vždycky, o Středulovo ego, o předvánoční nákupy odborářů v Praze (viz ten borec, co "se má dobře, ale stávkuje pro učitele...") a o trošku toho protisystémového bahýnka, které by Středula rád vytěžil.

@marek_rosa @MikeK1843773 @ylecun @TEDchris @vkhosla If implicit prompt would be enough, or explicit changes to the state of world would need to be added in the prompt.

@marek_rosa @MikeK1843773 @ylecun @TEDchris @vkhosla That might be interesting experiement to see how well can LLM consider applying the first half of the solutions and the impact on the world model when solving the next subproblem.

Fair if you define it in a self fulling way, but I define it as beating the median TED audience member in a "reasoning version of the Turing test". Maybe @TEDchris can run this test by randomly picking a few audience members at TED2024 and TED 2025 vs. GPTx... that I think is achievable though I also agree other methodologies like symbolic logic may make the process better or more efficient

@marek_rosa @MikeK1843773 @ylecun @TEDchris @vkhosla Somehow this feels too linear approach or we would need additional planner on top of the LLM that would be used to provide the prompts to navigate the problem solving. The challenge would be optimising for dependencies between subproblem outputs.

Indeed, it's difficult to bring this into effect. So, what's lacking in today's LLMs? Here's my perspective: 1) If LLMs could perfectly follow all instructions within a conversation (a sequence of prompts), aligning their behaviors accordingly, 2) If the length of the conversation (context window) could be extended indefinitely, Then, I believe, this multistep inference would become feasible, allowing us to achieve reasoning and planning over arbitrary horizons.