ceo @platzi 💚 7M students // board @gitlab foundation // yc w15

San Francisco, CA
I run an LLM offline on my laptop. A heavily quantized Qwen3.6, using LMStudio. MBP M4 Max 36GB RAM. It does 8-12 tokens per second. The context window is quite small (like 4000 tokens), but it's good enough for quick science/tech questions or helping me do a high-level grammar and narrative check while writing. I use it on flights with no wifi. It eats a LOT of battery life and gives me an eerie scifi feeling. I can tap all human knowledge, with no internet connection, in my "above average" laptop, compressed into 20GB. It's as easy as installing LMStudio and downloading the weights from HuggingFace.
47
37
635
35,772
You can use @MLXHub and outsource that memory to your iPhone or iPad 🫡
1
1
181
I should try
1
158
Replying to @freddier
20GB is a lossy sketch of human knowledge, and you stacked heavy quantization on top of it. I'd double check any science answer you plan to use after landing.
1
214
Still impressive
123
A 4,000-token context is enough for quick help, but it makes the laptop LLM a handy sidekick rather than a stand-in for a full research workflow.
I run an LLM offline on my laptop. A heavily quantized Qwen3.6, using LMStudio. MBP M4 Max 36GB RAM. It does 8-12 tokens per second. The context window is quite small (like 4000 tokens), but it's good enough for quick science/tech questions or helping me do a high-level grammar and narrative check while writing. I use it on flights with no wifi. It eats a LOT of battery life and gives me an eerie scifi feeling. I can tap all human knowledge, with no internet connection, in my "above average" laptop, compressed into 20GB. It's as easy as installing LMStudio and downloading the weights from HuggingFace.
1
1
1,094
It’s also helpful in winter to heat yourself up
2
986
Replying to @freddier
4000 tokens gets tight fast
1
1,887
For all human knowledge in a laptop in the sky, I'm pretty impressed.
3
15
1,754
Incentive misalignment. Never show the deck first thing.
Guys. Please. Go through the deck. I cannot emphasize enough how much it decreases the chances of getting an investor over the line when you decide to "just chat" with no visuals. How much time do you spend prepping for a raise? Between data room, deck design, narrative creation, etc? Usually 50-100 hrs? Why in the world would you then, at the ten yard line of a fundraising process (the weeks when you are actually pitching), decide to go about things in a way that yields lower levels of communication efficiency and information transfer? Especially if it is a pre-seed/seed/A raise where things are still very much being presented in the absence of empirical proof or data. Maybe 1-2% of the population in tech/startups is capable of running a 30-60 minute meeting with absolutely no materials and captivating a room with nothing but their own voice. For everyone else, your materials are a lever used to drive people towards the outcome you seek. Not using all tools available to you is silly! I promise you, whatever downside you think may come from presenting a pitch deck - coming across as robotic, rehearsed, so on and so forth - pales in comparison to the downside of not being able to control the conversation.
2
1
62
15,179
If one were excited and confident about what is being presented, what's the incentive misalignment? Isn't it in your best interest to be done fundraising ASAP so you can focus on building the company?
2
5
982
That’d imply that presenting the deck asap is directly correlated to faster financing.
1
3
417
Pregunta a ChatGPT: Vísteme con el traje tradicional del país que parece que soy: Y mega acertó 🫰🏼🇨🇴
5
39
3,979
heey
1
2
585
Claude Code auto classifier seems more strict now and it's becoming really hard to automate work with anything related to money and Opus 5.5 Classic Anthropic, goddammit.
12
3
195
19,894
Freddy Vega retweeted
"Arriésguense y lancen algo nuevo y grande". Freddy Vega, CEO y cofundador de @Platzi, nos deja un fuerte llamado a pensar en grande. La era de la inteligencia artificial es una oportunidad inmensa que nuestras empresas y el país no pueden dejar pasar. #INC2026 #Competitividad #20AñosEsMucho
1
6
16
1,136
Replying to @freddier
Genial tu conferencia en CPC - INC 2026 - 2027.
1
2
1,130
No sleep tonight girls and boys.
Introducing Gemini 4 Argon – our new frontier model. It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.
36
28
751
65,707
Noooo running on empty already
1
1
996
Keeps going
2
877
Freddy Vega retweeted
Claude Code course 💚 Live, in streaming, tomorrow for all Platzi students.
1
3
52
7,600
My 166 month old will be in the local news Friday. I will live tweet it, it’ll be THE event of MY year. So PROUD 🥹
9
73
4,269
Love everything about this tweet
1
1
429
Mr Robot is a very important TV show. A visual museum of what hackers were like before LLMs.
22
95
1,222
47,243
Learning from the best teacher @freddier at @platzi. Desde El Salvador 🇸🇻
1
4
452
Sonnet 5.5 Medium seems excellent and fast. At High, even with 30%+ token output speed, it writes more tokens than Opus and ends up less efficient. If you're using Sonnet 5.5 above Medium, just switch those tasks to Opus 5.5
7
10
314
17,173
Medium that beats Opus on cost-per-done is the whole story. Max intelligence is a demo. Throughput per dollar is the business.
1
821
For Sonnet sure. For Opus, Max has interesting use in CC
663