Carsten Ullrich

@OfficialLoganK ullrichc.github.io/stahnsdorf/info A web app for the most beautiful cemetery of Europe: Kirchhof Stahnsdorf, close to Berlin. Positions of interest, who is buried where, collections, ... Unaffordable for the institution to buy such an app, two days of work with Antigravity

we are designing something special for Google IO and we want you to be part of it reply with your AI Studio app, along with a 1-sentence story on how and why you vibe coded it

@yapper_so yapper

@aakashgupta @grok what is the source of this article?

The creator of Claude Code just showed you what an early AGI workflow actually looks like. Fifteen parallel instances, each functioning as an autonomous worker executing tasks unsupervised for hours. Starting sessions from your phone in the morning, checking output later in the day. Delegation to an entity you trust to figure things out while you’re not watching. The CLAUDE.md system reveals the real shift. Tagging @.claude on coworkers’ PRs to add learnings treats the AI like an onboarding engineer who accumulates institutional knowledge. You maintain a mistakes file for someone who learns. 2.5k tokens of accumulated corrections functioning as memory for an entity that would otherwise forget everything between sessions. Every sprint, the AI gets smarter within that codebase. Anthropic’s internal data shows 27% of Claude-assisted work “wouldn’t have been done otherwise.” Engineers building interactive dashboards, writing exploratory code, creating tools for small fixes they would have skipped before. The productivity gain extends beyond speed on existing tasks into expanding what gets attempted at all. The “manager of AI agents” framing keeps appearing because that’s genuinely the job now. One Anthropic engineer estimated their work shifted “70%+ to being a code reviewer/reviser rather than a net-new code writer.” Another described “taking accountability for the work of 1, 5, or 100 Claudes” as their future role. When Anthropic doubled their engineering headcount, PR throughput increased 67% because Claude Code handled the implementation load. The current bottleneck is reliability. Give Claude a browser to test its own UI and quality jumps 2-3x. Give it a test suite and it self-corrects. The model already possesses the capability. Building systems that let you trust output without reviewing every line unlocks the rest. Claude Code wrote 80% of its own codebase. Last year Anthropic engineers reported Claude helped with 30% of work. This year it’s 60%. The slope is steep. Commanding autonomous units instead of typing syntax. Directing intelligence rather than performing labor. That’s the feeling developers describe when this workflow clicks. Plan mode for every task. CLAUDE.md for accumulated learnings. Verification loops for trust. Slash commands for repetitive workflows. The setup is surprisingly vanilla. The power compounds through treating AI capability as infrastructure that improves over time. Fifteen parallel sessions supervised by one human brain, each trusted to execute independently, producing working code that ships to production. That’s the shape of what’s coming.

@Melt_Dem No Scifi-list is not complete without Ursula Le Guin

finished re-reading ender’s game and i forgot how phenomenal it was continue to believe reading sci fi is the greatest thing anyone interested in understanding, shaping, and directing the many manifolds of humanity’s future can and should do my sci fi reading list 👇

The hidden links between a giant of investigative journalism (e.g. ⁦@derspiegel)⁩ and the US government. 🚨 mediapart.fr/en/journal/int…

@NCseventeen Is this the Snyder cut?

The holy grail of workprints "Apocalypse Now" 5 hour Cannes Cut (Workprint) Download Link is in comments

AI can predict public opinion by modeling the media diets of population subgroups. By fine-tuning language models like BERT on news & media consumed by specific groups, researchers could accurately predict their survey responses on issues like COVID-19. arxiv.org/pdf/2303.16779…

@Rainmaker1973 Google Gemini says: Sure, I can help you solve the puzzle. There are 12 triangles containing only one black circle. .

I'm gone for 12 hours and OpenAI just tossed the whole world upside down again 🤯 Even though I'm keeping up with all the Generative AI progress, Sora is such a big jump in video generation capabilities that it feels unreal. Time for a mega thread 🔽

Example of LLM coupled with action: 'Coscientist,' a GPT-4 powered AI for autonomous chemical research. It merges internet search, coding, and automation to expedite experiments, like reaction optimization nature.com/articles/s4158…

AI (ChatGPT, DALL-E) has lower CO2 emissions than humans in writing/illustrating. ChatGPT emits 1100x less CO2 per text page than a US resident, DALL-E 2500x less. Intriguing insight into AI's environmental impact. arxiv.org/abs/2303.06219

Prompt engineering vs. model fine-tuning: A new study reveals GPT-4 outperforms specialized models like BioGPT & Med-PaLM in medical benchmarks using clever prompting. Shows the power of 'generic' models with the right prompts. doi.org/10.48550/arXiv…

@fchollet With X, do you mean what was previously known as Twitter? Seriously, I was quite confused for a few seconds :-)

To understand X means you have the ability to act appropriately in response to situations related to X -- for instance, you understand how to make coffee in a kitchen if you can walk into a random kitchen and make coffee.

Very technical & impressive visualization and explanation of how large language models compute their answers. Worth clicking through the overview (using nano-gpt), which by itself is already impressive, & then on GPT-3 to see the difference in scale. bbycroft.net/llm #llm

That is impressive. Looking forward to get my hands on it.

Separating quantum computing hype from reality is tough. Prof. Juliane Krämer, a cryptography expert, estimates 20-50 years for significant practical application. Big physics challenges remain (source in German) inf.gi.de/03/wann-wenn-n…

Commonwealth Bank of Australia explores generative AI for 'synthetic customers' to test new products. Sparse on details. My guess: defining customer groups, detailing them, then asking an AI their take on new offerings. finextra.com/newsarticle/43… #Banking #generativeai

@markopolojarvi @HochreiterSepp @bifoldberlin GPT4 was 100 million, wasn't it? He says the 300 million are the total, and wants to increase model size step by step

@HochreiterSepp presented xLSTM at @bifoldberlin 's AI symposium, outperforming transformers like GPT on small datasets. Next: train on bigger data, but needs 300M€ & European support to keep the tech local. Would back it if I could! #generativeai

@newplatonism @MarioKrenn6240 @HochreiterSepp @bifoldberlin He claims his new approach has solved the scaling problem, but no publication yet

@EmaMazzi76 @JagersbergKnut @HochreiterSepp @bifoldberlin Nothing published yet...

@ullrich @JagersbergKnut @HochreiterSepp @bifoldberlin Interesting, where the comparison with the transformers can be seen?