Jorge De Corte

Exploring the world of AI and Machine Learning

[Jorge De Corte @JorgeDeCorte]

@elonmusk I’m unable to activate the fog lights on my Tesla, even in heavy fog. Since the automatic lighting system didn’t turn on the headlights, the fog light button remained disabled. As a driver, I want to have the ability to override the AI system. Would it be possible to allow manual activation of the fog lights at any time, even if the car thinks it is not necessary?

[Chris Anderson @TEDchris]

With a certain amount of trepidation, I'm posting this open letter to @elonmusk, someone I have admired, but who, right now, is causing me concern. I know I'm not alone in thinking these thoughts. Please like or repost if you're willing... And Elon, if you're listening, please know this is offered in a constructive spirit.

-----

Dear Elon,

A thought hit me this morning.

On top of all your achievements in technology and entrepreneurship, you have become this century's single most influential writer.  You have more than 200 million followers on a powerful platform that you yourself control. And the words that you write flow well beyond just those followers into almost every user of X and far beyond courtesy of extensive media coverage.   This makes Rupert Murdoch at his prime seem inconsequential.

It must feel exhilarating. You made a huge gamble buying Twitter and it seems to have paid off spectacularly to the point that you can use it to massively impact the world, including changing governments.  You believe that X can largely replace most mainstream media. It is the new platform for citizen journalism, and you are citizen number 1.  You don’t need the hassle of editors and fact-checkers. Every single thing you post garners millions of likes and reposts. In a heartbeat you can change the global conversation.  No one in history has had this much power.

So there’s a lot at stake here, and, as it happens, journalism is something I care deeply about. I began my career as a journalist because I believed that good journalism was essential for the healthy functioning of democracy. Today I am worried — quite deeply worried, actually — that in your triumphant seizing of the global conversation, some of the core tenets of journalism are being forgotten. Without them, I think your efforts to make X the respected home of citizen journalism will fail.

There are numerous journalistic principles that matter — Grok can summarize them quite nicely. But there’s one in particular that’s been troubling me. It’s the fairness doctrine. The one that says that before you publish savagely critical claims about an individual, or an institution, you reach out to them for their side of the story. After all, just possibly, you may have missed a key fact or two that would change how people assess what has happened. Just possibly your sources were motivated to cause damage to that individual. Just possibly there’s an alternative explanation of what happened.

So, for example, when you tell hundreds of millions of people that someone should be hanged or jailed for outrageous crimes against humanity, just possibly you should first sound out what those who know those people really well would say about them. Some of your recent posts could literally get someone killed. Do you really want to risk that? How is it possible that you can do this at the very same time that you’re calling on people to make X more positive, more beautiful?  You say you want to maximize un-regretted user-seconds on X.  By far the simplest way you could do this, Elon, is simply to thoughtfully edit what you yourself post.

I get that from your eyes the issues you are championing are unbelievably important and worthy of extreme efforts. But the way you are presenting them is not citizen journalism. It's playground bullying. It’s crass and it’s cruel, and it’s therefore not nearly as effective as it could be.  You’re hearing the cheers of your most loyal followers, but missing the fact you’re making yourself a laughing stock among many who you really want on your side. Long-term that’s going to damage X, your other businesses, and indeed your long-term dreams for humanity.  No one wants to follow a playground bully to Mars.

I miss the old Elon. You can be funny, interesting, insightful and inspiring. You've fought incredibly hard for what you've built. And you may feel you're entitled to do whatever the hell you want with it. But I also know that you understand the danger of holding too tightly to the ring of power, how it can distort someone's judgement and turn them ugly.

I’m hoping you can loosen that ring just a little. For the love of humanity that you profess, I really urge you to embrace the fairness doctrine and showcase a better face of X.

Thanks for listening.

Chris

[Jorge De Corte @JorgeDeCorte]

That was amazing🤯

[Andrew McCalip @andrewmccalip]

BOOSTER HAS LANDED!!!!!!

[Jorge De Corte @JorgeDeCorte]

@AIatMeta Great! Looking forward to the weights!

[Jorge De Corte @JorgeDeCorte]

@thenerd_be 🤣

[Jorge De Corte @JorgeDeCorte]

@mattshumer_ @_philschmid @GlaiveAI Congrats! Can’t wait to try it out

[Yann LeCun @ylecun]

#Openalwayswins

[Jorge De Corte @JorgeDeCorte]

@ylecun Great! Is Meta planning an API for the 405B model? It’s needs quite the infrastructure😆

[Jorge De Corte @JorgeDeCorte]

@Ahmed_Masry97 Wow thank you very much!

[Jorge De Corte @JorgeDeCorte]

@Ahmed_Masry97 Great model! Thx! When will you release the generated dataset?

[Jorge De Corte @JorgeDeCorte]

@gdb I hate that🫣

[Jorge De Corte @JorgeDeCorte]

@joristi Het is bijna te gemakkelijk😂

[Jorge De Corte @JorgeDeCorte]

@BramVanroy But even 50 conversations or question answers requires a lot of work!!

[Jorge De Corte @JorgeDeCorte]

@BramVanroy We are experimenting using RAGAS for our RAG specific applications. It requires a lot of work to setup the ground thruth but so far it seems to work. Our plan is to switch models or checkpoints and compare scores.

[Jorge De Corte @JorgeDeCorte]

@jos_van_josbeek @DOBBELAEREW Deze statistiek lijkt me een net een reden voor meer privacy

[Jorge De Corte @JorgeDeCorte]

@BramVanroy Probably benchmark (depends on what the benchmark is) a handful of conversations is not representative.

[Jorge De Corte @JorgeDeCorte]

@ClementDelangue Thanks!

[Jorge De Corte @JorgeDeCorte]

How you can control your own generative AI without a dependency on the American tech industry!

[ReBatch @rebatch_ml]

Using generative AI from services such as Azure OpenAI or Anthropic is not without risk. You are effectively outsourcing your (artificial) intelligence to a third party you do not know.

👉 What you can do about it: https://t.co/qzvxIpRsQ0

[Jorge De Corte @JorgeDeCorte]

I haven’t tested the model yet, but if these claims are true. It’s pretty impressive work. The only downside is you have to retrain when a new function is added.

[Philipp Schmid @_philschmid]

On-Device 2B LLMs for actions, outperform GPT-4 🤯 The “Octopus v2: On-device language model for super agent” proposes a new method to create on-device agents. 📱🔄

Implementation
1️⃣ Define supported functions as special tokens, e.g. <func_1> and add them to the tokenizer
2️⃣ Collect/Generate 100-1,000 data points for each function set, e.g. “take a photo” → <func2>
3️⃣ Fine-tune LLM with updated tokens on dataset to learn when to use which function and what parameters to use, here Gemma 2B
4️⃣ Deploy the trained and fine-tuned model on edge devices or build agents

Insights
🔑 Adds 1 special token per "function" to reduce error rate which functions to use
✅ 98.095% accuracy with just 100 samples per function.
⚡ 35x faster latency than Llama-7B with a RAG-based function calling
⏱️ Works on-device within 1.1 to 1.7 seconds for typical queries
💸 Synthetic data generation for one new task ~$0.0224 for 1,000 samples
⛓️ Supports nested, and chained function calls
🤖 Could be used in smart devices, like Alexa, to interact with maps, food delivery etc.
❓ Unsure if this can be scaled to hundreds of different functions, currently trained on 20.

Paper: https://t.co/1rowxFQnNc
Models: https://t.co/gzxfKRv5bJ

Interesting idea to incorporate the functions into the model with fine-tuning to get reliable generation from small LLMs. But this reduces the flexibility on what function can be used. Do you think we need “agents” that can use any function, or are ~20-100 enough? 🤔

[Jorge De Corte @JorgeDeCorte]

@_philschmid I bet Apple is reading this 👀