Infinitech

Working in AI to deliver a next-gen, multi-channel collaborative ERP where LLMs dynamically manage the workflow.

[Infinitech @Infinitech_it]

@DenLoginoff @ivanfioravanti Intelligence-wise, they seem the same to me; they choose more or less the same patterns, make the same mistakes, and overlook the same things. When the task is complex, you have to guide them. GLM seems more mature to me. However, the cost is 10 times less and can't be ignored.

[Infinitech @Infinitech_it]

@DenLoginoff @ivanfioravanti Compared to GLM-5.2, Deepseek is a little less tolerant of prompt configuration. If you enable thinking, the JSON response often drops some parts. If you change the temperature, etc., it's the same. I'm trying to make my agent tolerant, but is a good model. I think I'll use it.

[Infinitech @Infinitech_it]

@DenLoginoff @ivanfioravanti I don't want compare on the same task but in a long task building an ERP with a codebase.

[Infinitech @Infinitech_it]

@DenLoginoff @ivanfioravanti T9

[Infinitech @Infinitech_it]

@DenLoginoff @ivanfioravanti This night or tomorrow I'll try with low temperature and think high

[Infinitech @Infinitech_it]

@DenLoginoff @ivanfioravanti I'm doubtful, with temperature 1 (which in my opinion is wrong for an agent, but recommended by deepseek) the agent has too many liberties and doesn't follow instructions very well and doesn't seem very intelligent to me (with think=low)

[Infinitech @Infinitech_it]

@ivanfioravanti At the moment it seems in line with glm-5.2 I want to try with different temperatures because it seems much more easy-going than GLM.

[Infinitech @Infinitech_it]

@ivanfioravanti No, no but you should try this AI agent https://t.co/BWfZ8do6kS

[Infinitech @Infinitech_it]

Try it with our AI agent and then tell us the difference. https://t.co/BWfZ8do6kS

[Infinitech @Infinitech_it]

We're online with our AI agent https://t.co/BWfZ8do6kS — don't tell anyone. If you want to stop programming and destroy the old world, this is the place to be. Keep it just for yourself. PS: If you contact me for feedback after the free trial you will get an extra large discount.

[Infinitech @Infinitech_it]

@adamkjohnston Among these immigrants, the majority are student visas, Ukrainians, and workers. There hasn't been mass immigration, but rather a shortage of labor force has been addressed with stringent regulations, especially in recent years.

[Infinitech @Infinitech_it]

@ivanfioravanti I personally hired a young guy to follow my instructions, ask AI ​​for explanations, ask for details, and analyze the implementation. After that, I'll do the same for the entire development process, but with a higher level of abstraction. How many expert programmers do you need?

[Infinitech @Infinitech_it]

@ivanfioravanti So far, I've developed a context-optimized AI agent that saves me 70% of the tokens and performs at the top of the AI ​​agents available. Combined with a Chinese LLM that costs 10 times less, I'd say I spend 1/30th of what the average programmer spends.

[Infinitech @Infinitech_it]

@pupposandro I don't know what i could run...but i'll see

[Infinitech @Infinitech_it]

@pupposandro 2billion of token for month with cinese llm

[Infinitech @Infinitech_it]

@ivanfioravanti @antirez As for the backend, I agree, but for the frontend, a framework is a lifesaver... I personally updated a 20-year-old PHP/JQuery/Ajax framework. Two-way communication, modern CSS, fast and lightweight, and I also like some trends like "svelte"—AI understands simple things instantly

[Infinitech @Infinitech_it]

@ivanfioravanti I also create a swarm of agents, with a central agent that download the issue from GitHub and send the work to the right agent and project. I can control all the swarm via whatsapp. It's as having ten employers. A Revolution

[Infinitech @Infinitech_it]

@ivanfioravanti Yes....with forgecode I don't write code anymore. It understand your codebase and It write in your style. It's super, but i don't know if it's for the agent or for the model. The agent work quite well also with smaller model.