Ezequiel Sculli (@ezesculli) header
Ezequiel Sculli (@ezesculli) profile picture

Founder & CPO/CTO at Darwin AI (https://nitter.cf/t.co/nJcdtHZvEr)

Joined August 2008
Muy orgulloso de compartirles el lanzamiento de Darwin AI (getdarwin.ai), junto con el anuncio de nuestra ronda Seed de 2.5MM USD. 🚀🚀🚀 Estoy muy feliz de encarar este desafio junto a @lautaschiaffino y muchas personas más que confiaron en nosotros desde el inicio
10
6
2
82
6,287
Se me acabarón los tokens de Claude, realmente no tiene sentido trabajar sin super poderes. Mejor descansar un rato y volver cuando se reseteé la sesión o pagar un plan extra 🤑
2
151
OpenAI Symphony: Es un proyecto open-source de OpenAI que orquestra multiples Codex resolviendo tareas en un Board de Linear. Lo interesante es que ya no solo te dan el codigo completo de la implementación, sino la definición exaustiva para que lo puedas implementar vos 🧵
1
82
Podes implementarlo vos directamente en tu Claude Code, Codex o Cursor: github.com/openai/symphony/b… Puede ser que el futuro del open-source sea no solo tener una implementación, sino un PRD (product requirements) de como implementar la solución vos mismo.
1
72
El 70-80% de las empresas de software no van a sobrevivir la transición a AI. No porque la tecnología no esté lista, sino porque no van a poder cambiar su forma de trabajar lo suficientemente rápido 🤯🤯🤯
54
Los mejores Devs van a ser los que mejor sepan hablar con clientes. Y los mejores product managers van a ser los que mejor entiendan el código a fondo. AI está borrando la línea entre "Producto" y "Desarrollo". Los perfiles híbridos son los que mejor se van a adaptar 🚀🚀
3
89
Ezequiel Sculli retweeted
Meet Codex CLI—an open-source local coding agent that turns natural language into working code. Tell Codex CLI what to build, fix, or explain, then watch it bring your ideas to life.
99
447
195
3,011
647,028
Toda la humanidad intentando escapar de la gravedad que la tuvo confinada durante 50,000 años
Once again, “Viva La Starship en Mexico!!” 🇲🇽 🙌🤣🚀
1
306
Ezequiel Sculli retweeted
Starship reentering like a meteor
8,655
33,189
1,721
647,878
54,419,671
Ezequiel Sculli retweeted
Super Heavy has splashed down in the Gulf of Mexico
2,207
11,877
2,931
69,717
57,897,289
La fórmula secreta para construir productos AI-native 🤖
This week, Google announced a doubling of Gemini Pro 1.5's input context window from 1 million to 2 million tokens, and OpenAI released GPT-4o, which generates tokens 2x faster and 50% cheaper than GPT-4 Turbo and natively accepts and generates multimodal tokens. I view these developments as the latest in an 18-month trend. Given the improvements we've seen, best practices for developers have changed as well. Since the launch of ChatGPT in Nov 2022, with key milestones that include the releases of GPT-4, Gemini 1.5 Pro, Claude 3 Opus, and Llama 3-70B, many model providers have improved their capabilities in two important ways: (i) reasoning, which allows LLMs to think through complex concepts and and follow complex instructions; and (ii) longer input context windows. The reasoning capability of GPT-4 and other advanced models makes them quite good at interpreting complex prompts with detailed instructions. Many people are used to dashing off a quick, 1- to 2-sentence query to an LLM. In contrast, when building applications, I see sophisticated teams frequently writing prompts that might be 1 to 2 pages long (my teams call them “mega-prompts”) that provide complex instructions to specify in detail how we’d like an LLM to perform a task. I still see teams not going far enough in terms of writing detailed instructions. For an example of a moderately lengthy prompt, take a look at Claude 3’s system prompt. It’s detailed and gives clear guidance on how Claude should behave. This is a very different style of prompting than we typically use with LLMs’ web user interfaces, where we might dash off a quick query and, if the response is unsatisfactory, clarify what we want through repeated conversational turns with the chatbot. Further, the increasing length of input context windows has added another technique to the developer’s toolkit. GPT-3 kicked off a lot of research on few-shot in-context learning. For example, if you’re using an LLM for text classification, you might give a handful — say 1 to 5 examples — of text snippets and their class labels, so that it can use those examples to generalize to additional texts. However, with longer input context windows — GPT-4o accepts 128,000 input tokens, Claude 3 Opus 200,000 tokens, and Gemini 1.5 Pro 1 million tokens (2 million just announced in a limited preview) — LLMs aren’t limited to a handful of examples. With many-shot learning, developers can give dozens, even hundreds of examples in the prompt, and this works better than few-shot learning. When building complex workflows, I see developers getting good results with this process: - Write quick, simple prompts and see how it does. - Based on where the output falls short, flesh out the prompt iteratively. This often leads to a longer, more detailed, prompt, perhaps even a mega-prompt. - If that’s still insufficient, consider few-shot or many-shot learning (if applicable) or, less frequently, fine-tuning. - If that still doesn’t yield the results you need, break down the task into subtasks and apply an agentic workflow. I hope a process like this will help you build applications more easily. If you’re interested in taking a deeper dive into prompting strategies, I recommend Microsoft's Medprompt paper (Nori et al., 2023), which lays out a complex set of prompting strategies that can lead to very good results. [Original text (with links): deeplearning.ai/the-batch/is… ]
3
254
Google está intentando resistir como pueda contra OpenAI! La competencia va a ser feroz!!!
One more day until #GoogleIO! We’re feeling 🤩. See you tomorrow for the latest news about AI, Search and more.
1
258
Se viene un mundo muy diferente! Es increible la capacidad de adaptación de la voz al contexto 🤯🤯🤯
Say hello to GPT-4o, our new flagship model which can reason across audio, vision, and text in real time: openai.com/index/hello-gpt-4… Text and image input rolling out today in API and ChatGPT with voice and video in the coming weeks.
3
302
Ezequiel Sculli retweeted
I recommend investing in Argentina
Great meeting with President ⁦@JMilei⁩!
19,690
57,583
11,499
509,185
132,776,902
Meta acaba de invertir mas de 30 Billones de USDs en chips e infraestructura para entrenar modelos de AI y LLMs. Es exactamente el mismo presupuesto que tuvo la NASA para llegar a la Luna 🤯🤯
6
226
Se invierte la ecuación? Human Workers trabajando para AI agents!
Excited to introduce a new project I've been working on called Payman! Payman is an AI Agent tool that gives Agents the ability to pay people for tasks they cannot do themselves. While many people imagine a future where humans pay AI agents for services they want completed, I believe that as AI agents become more advanced, they will be paying humans for tasks they can’t do. There will always be important roles for humans, and as we move towards an agent-driven world, Payman’s goal is to support a symbiotic relationship between AI agents and humans. Payman addresses three major challenges to make this collaboration possible: Access to Funds: AI agents can't open bank accounts due to current regulations. It might be a long time before this changes, if ever. Payman simplifies this by allowing AI agents with access to their own funds to spend as they want, without a bank account. Quality Task Completion: It’s hard for AI agents to find reliable, skilled human workers. While platforms like Fiverr and Upwork exist, they don’t meet the fast-paced and quality-specific needs of AI workflows. Payman is developing the largest vetted database of skilled workers that AI agents can tap into for task completions. Verification of Work: Ensuring that tasks are completed correctly is crucial. Payman is creating a suite of verification agents that will check that work meets task requirements, helping AI agents achieve their goals and ensuring humans are paid fairly. There are tons of use cases that Payman opens up for Agents! Design: Humans add creative input to help Agent's design better products. Code: Humans perform code reviews to ensure it meets specifications. Law: Humans provide insights to gauge public sentiment about legal cases so Agents can make better strategies. Gaming: Agents pay humans to complete real-world tasks in games. Medical: Medical professionals help to improve diagnostic accuracy for Agents. Sales: Humans execute sales strategies developed by AI agents. Marketing: Humans are hired to promote products based on the Agent's strategy. Right now this is still in early beta and I am looking for any Agent builders that are interested in adding superpowers to what their Agent can do! DM me if you’d like access or sign-up to the waitlist at paymanai.com/. If you’re interested in the project and want to help contribute, please send me over a DM! I’m looking for people passionate about the intersection of Humans and Agent’s working together.
190
Rewind ahora es Limitless!
Introducing Limitless: a personalized AI powered by what you’ve seen, said, or heard. It’s a web app, Mac app, Windows app, and a wearable. 0:06 Reveal 0:48 Why Limitless? 1:39 Demo 3:05 Pendant 4:27 Privacy 5:23 Confidential Cloud 6:36 Rewind 7:12 Roadmap 9:25 Vision
175
Creo que el desafío con AI es el mismo: reducir la fricción y complejidad para que más negocios puedan implementarlo lo más rápido posible!! 🚀🚀
Tobi Lutke explains what the VCs who passed on Shopify got wrong Tobi recounts pitching Shopify to VCs on Sand Hill Road a few years after founding Shopify. Investors passed because they thought the addressable market was too small. At the time, there were about 40,000-50,000 online stores, and even if Shopify captured 50% of the market, that still wouldn’t be a venture-scale business. When Tobi ran into the VC partner a few years ago, the partner asked Tobi what he missed (Shopify is valued at almost $100 billion today). Tobi explained: “You were actually correct, but what you didn’t realize was that Shopify was the solution to the very problem you identified. The reason there was only 40,000 online stores was because it was hard, expensive, and everyone who tried ran into all these brick walls of complexity, which Shopify, one after another, smoothed over and made simple to do.” Tobi believes this is a common mistake: “What a lot of free-market thinkers don’t understand is that between the demand and eventual supply lies friction. And I actually think that friction is probably the most potent force for shaping the planet that people just generally do not acknowledge… That was my theory when I turned my snowboard store into Shopify: there was a lot more people like me except there was too much friction which we needed to solve. And Shopify has proven out that every time we make the process simpler, there’s more consumption. At this point, we have a million merchants on Shopify, which is a mind-blowing number. So friction is a major component, and it’s something that software is uniquely good at reducing.” Video Source: @danmartell
2
252
Muchas gracias @diegotauziet por invitarme a hablar de AI y de Darwin AI en la @UdeSA 🙌👏 Fue verdaderamente un privilegio tener está charla al lado de la ex-Gobernadora @mariuvidal PD: Ella estaba literalmente en el aula de al lado 😝 seguramente algo escucho de fondo
2
5
472
Ezequiel Sculli retweeted
Now that we have about two dozen LLMs in the market, here are the dimensions that matter when it comes to using them. Reasoning - Claude 3 Opus beats everything out Code - GPT-4 is still king here Cost - Claude Haiku is your best bet Latency - Claude or a local open-source model is worse best Fine-Tuning - If you must fine-tune, I would vote for Mistral. Best Local Model - If your security team throws a fit for no good reason and insists on a local model - Qwen 72B or Smaug-2 (fine-tune on Qwen). Qwen 72B instruct is on top of the human eval leaderboard. Best small local model - Starling-7b. Again, on top of that leaderboard. I purposely didn't include the extended context here as it has yet to translate to good context understanding. For now, I prefer sticking to 128K and dealing with it.
47
163
13
774
140,889