Alvin Ariesta
@ariesta.id
Looking for PhD opportunity: AI for the environment, AI for science.
https://ariesta.id
https://linkedin.com/in/alvinariesta/
https://ariesta.id
https://linkedin.com/in/alvinariesta/
Reposted by Alvin Ariesta
Don't leave AI to the STEM folks.
They are often far worse at getting AI to do stuff than those with a liberal arts or social science bent. LLMs are built from the vast corpus human expression, and knowing the history & obscure corners of human works lets you do far more with AI & get its limits.
They are often far worse at getting AI to do stuff than those with a liberal arts or social science bent. LLMs are built from the vast corpus human expression, and knowing the history & obscure corners of human works lets you do far more with AI & get its limits.
July 20, 2025 at 6:06 PM
Don't leave AI to the STEM folks.
They are often far worse at getting AI to do stuff than those with a liberal arts or social science bent. LLMs are built from the vast corpus human expression, and knowing the history & obscure corners of human works lets you do far more with AI & get its limits.
They are often far worse at getting AI to do stuff than those with a liberal arts or social science bent. LLMs are built from the vast corpus human expression, and knowing the history & obscure corners of human works lets you do far more with AI & get its limits.
My humble definition of "big data"
June 29, 2025 at 8:52 AM
My humble definition of "big data"
Reflection:
marimo is not a "notebook". It's a "published book", presentable and reliable to be shared. Jupyter is still the best notebook, especially on Google Colab.
Why? For me just 1 problem in marimo: unable to redefine variables without some code gymnastics (docs.marimo.io/guides/under...)
marimo is not a "notebook". It's a "published book", presentable and reliable to be shared. Jupyter is still the best notebook, especially on Google Colab.
Why? For me just 1 problem in marimo: unable to redefine variables without some code gymnastics (docs.marimo.io/guides/under...)
June 9, 2025 at 3:59 PM
Reflection:
marimo is not a "notebook". It's a "published book", presentable and reliable to be shared. Jupyter is still the best notebook, especially on Google Colab.
Why? For me just 1 problem in marimo: unable to redefine variables without some code gymnastics (docs.marimo.io/guides/under...)
marimo is not a "notebook". It's a "published book", presentable and reliable to be shared. Jupyter is still the best notebook, especially on Google Colab.
Why? For me just 1 problem in marimo: unable to redefine variables without some code gymnastics (docs.marimo.io/guides/under...)
Boy... cloud parallelism is satisfying :chefkiss:
June 9, 2025 at 3:52 PM
Boy... cloud parallelism is satisfying :chefkiss:
June 9, 2025 at 1:39 PM
Reposted by Alvin Ariesta
Comma v0.1 1T and 2T-7B are two brand new LLMs trained exclusively on public domain and openly licensed text!
I got the 2T one running locally on my Mac after porting it to MLX - notes on that here: simonwillison.net/2025/Jun/7/c...
I got the 2T one running locally on my Mac after porting it to MLX - notes on that here: simonwillison.net/2025/Jun/7/c...
Comma v0.1 1T and 2T—7B LLMs trained on openly licensed text
It’s been a long time coming, but we finally have some promising LLMs to try out which are trained entirely on openly licensed text! EleutherAI released the Pile four and …
simonwillison.net
June 8, 2025 at 12:10 AM
Comma v0.1 1T and 2T-7B are two brand new LLMs trained exclusively on public domain and openly licensed text!
I got the 2T one running locally on my Mac after porting it to MLX - notes on that here: simonwillison.net/2025/Jun/7/c...
I got the 2T one running locally on my Mac after porting it to MLX - notes on that here: simonwillison.net/2025/Jun/7/c...
Reposted by Alvin Ariesta
This is a good overview of AI power use (small at individual level, big in aggregate).
One thing that struck me: they tested LLama 3.1 405B and it averaged 3,353 joules per prompt. That is the equivalent of 2 minutes 50 seconds of human brain activity. www.technologyreview.com/2025/05/20/1...
One thing that struck me: they tested LLama 3.1 405B and it averaged 3,353 joules per prompt. That is the equivalent of 2 minutes 50 seconds of human brain activity. www.technologyreview.com/2025/05/20/1...
We did the math on AI’s energy footprint. Here’s the story you haven’t heard.
The emissions from individual AI text, image, and video queries seem small—until you add up what the industry isn’t tracking and consider where it’s heading next.
www.technologyreview.com
May 21, 2025 at 1:16 PM
This is a good overview of AI power use (small at individual level, big in aggregate).
One thing that struck me: they tested LLama 3.1 405B and it averaged 3,353 joules per prompt. That is the equivalent of 2 minutes 50 seconds of human brain activity. www.technologyreview.com/2025/05/20/1...
One thing that struck me: they tested LLama 3.1 405B and it averaged 3,353 joules per prompt. That is the equivalent of 2 minutes 50 seconds of human brain activity. www.technologyreview.com/2025/05/20/1...
Thank you DeepSeek. "Worry time", a new term for me.
March 14, 2025 at 4:28 AM
Thank you DeepSeek. "Worry time", a new term for me.
Most people don't set reasonable expectations.
I usually forgot to set expectations too, after not interacting with LLM for a while
I usually forgot to set expectations too, after not interacting with LLM for a while
Here's the table of contents for my lengthy new piece on how I use LLMs to help me write code simonwillison.net/2025/Mar/11/...
March 14, 2025 at 1:43 AM
Most people don't set reasonable expectations.
I usually forgot to set expectations too, after not interacting with LLM for a while
I usually forgot to set expectations too, after not interacting with LLM for a while
It seems that OpenAI Deep Research is so useful because of its search capability. I'm not really impressed with OpenAI's high-end models outside Deep Research.
Imagine if they are truly open and people can connect the search indexing with their own source priority list
Imagine if they are truly open and people can connect the search indexing with their own source priority list
February 12, 2025 at 3:32 PM
It seems that OpenAI Deep Research is so useful because of its search capability. I'm not really impressed with OpenAI's high-end models outside Deep Research.
Imagine if they are truly open and people can connect the search indexing with their own source priority list
Imagine if they are truly open and people can connect the search indexing with their own source priority list
My current LLM choice:
- large context, image, video: Gemini via AI Studio
- code: Claude (web)
- search: DeepSeek web, Perplexity
- writing: Llama (GPU rent -- just because I have unused credits)
- large context, image, video: Gemini via AI Studio
- code: Claude (web)
- search: DeepSeek web, Perplexity
- writing: Llama (GPU rent -- just because I have unused credits)
January 28, 2025 at 6:03 PM
My current LLM choice:
- large context, image, video: Gemini via AI Studio
- code: Claude (web)
- search: DeepSeek web, Perplexity
- writing: Llama (GPU rent -- just because I have unused credits)
- large context, image, video: Gemini via AI Studio
- code: Claude (web)
- search: DeepSeek web, Perplexity
- writing: Llama (GPU rent -- just because I have unused credits)
Reposted by Alvin Ariesta
Anthropic co-founder Jack Clark muses on whether AI systems will give an unfair advantage to people with a fiercely curious nature
importai.substack.com/p/import-ai-...
importai.substack.com/p/import-ai-...
![[…] in the era where these AI systems are true 'everything machines', people will out-compete one another by being increasingly bold and agentic (pun intended!) in how they use these systems, rather than in developing specific technical skills to interface with the systems.
We should all intuitively understand that none of this will be fair. Curiosity and the mindset of being curious and trying a lot of stuff is neither evenly distributed or generally nurtured. Therefore, I'm coming around to the idea that one of the greatest risks lying ahead of us will be the social disruptions that arrive when the new winners of the AI revolution are made - and the winners will be those people who have exercised a whole bunch of curiosity with the AI systems available to them.
- Jack Clark](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:kft6lu4trxowqmter2b6vg6z/bafkreiai5uemf27nso53xg7pdkf43ifz2cnid35phk3twpcsj7pbgo3ogy@jpeg)
January 28, 2025 at 2:45 AM
Anthropic co-founder Jack Clark muses on whether AI systems will give an unfair advantage to people with a fiercely curious nature
importai.substack.com/p/import-ai-...
importai.substack.com/p/import-ai-...
Yeah, if you don't like "AI", at least criticize it productively
Outrightly dismissing it is not productive
Outrightly dismissing it is not productive
I worry people are not reckoning with what AI can and can't do well. I post about it occasionally, and there are always people who insist that it's just scraping and regurgitating, or even that there's an actual human writing the responses.
January 25, 2025 at 12:50 AM
Yeah, if you don't like "AI", at least criticize it productively
Outrightly dismissing it is not productive
Outrightly dismissing it is not productive
DeepSeek just released R1, an open weight "thinking" model like OpenAI's O1. The difference is:
- people said O1's better
- O1 is closed, R1 is open weight
- R1 is MIT-licensed, can use it to train/finetune other models
- cannot see O1 thoughts, can see R1's
api-docs.deepseek.com/news/news250...
- people said O1's better
- O1 is closed, R1 is open weight
- R1 is MIT-licensed, can use it to train/finetune other models
- cannot see O1 thoughts, can see R1's
api-docs.deepseek.com/news/news250...
DeepSeek-R1 Release | DeepSeek API Docs
* ⚡ Performance on par with OpenAI-o1
api-docs.deepseek.com
January 21, 2025 at 10:24 AM
DeepSeek just released R1, an open weight "thinking" model like OpenAI's O1. The difference is:
- people said O1's better
- O1 is closed, R1 is open weight
- R1 is MIT-licensed, can use it to train/finetune other models
- cannot see O1 thoughts, can see R1's
api-docs.deepseek.com/news/news250...
- people said O1's better
- O1 is closed, R1 is open weight
- R1 is MIT-licensed, can use it to train/finetune other models
- cannot see O1 thoughts, can see R1's
api-docs.deepseek.com/news/news250...
Come on Bluesky... If I don't keep the other-side account, I wouldn't find out about DeepSeek R1 open weight release. Even my LinkedIn timeline is better.
January 21, 2025 at 10:21 AM
Come on Bluesky... If I don't keep the other-side account, I wouldn't find out about DeepSeek R1 open weight release. Even my LinkedIn timeline is better.
I decided to dedicate more time to my master's research and resigned from a full-time job
I hoped to get some time off from reading emails, but then I made the mistake of turning on Google Scholar alerts for the researchers I follow
I hoped to get some time off from reading emails, but then I made the mistake of turning on Google Scholar alerts for the researchers I follow
January 19, 2025 at 9:43 AM
I decided to dedicate more time to my master's research and resigned from a full-time job
I hoped to get some time off from reading emails, but then I made the mistake of turning on Google Scholar alerts for the researchers I follow
I hoped to get some time off from reading emails, but then I made the mistake of turning on Google Scholar alerts for the researchers I follow
Reposted by Alvin Ariesta
Does ChatGPT use 10x more energy than a standard Google search? https://engineeringprompts.substack.com/p/does-chatgpt-use-10x-more-energy #AI #climate
January 17, 2025 at 2:38 PM
Does ChatGPT use 10x more energy than a standard Google search? https://engineeringprompts.substack.com/p/does-chatgpt-use-10x-more-energy #AI #climate
Markdown Is All You Need
January 8, 2025 at 11:54 AM
Markdown Is All You Need
Blessing in disguise from the WordPress drama. Maintaining a blog with MkDocs, Docker, and Coolify is a lot simpler.
Learning Docker is a big hurdle though. New post:
ariesta.id/blog/2025/01...
Learning Docker is a big hurdle though. New post:
ariesta.id/blog/2025/01...
Blog built by mkdocs-material, served by NGINX inside a Docker container - ariesta.id
ariesta.id
January 8, 2025 at 11:43 AM
Blessing in disguise from the WordPress drama. Maintaining a blog with MkDocs, Docker, and Coolify is a lot simpler.
Learning Docker is a big hurdle though. New post:
ariesta.id/blog/2025/01...
Learning Docker is a big hurdle though. New post:
ariesta.id/blog/2025/01...
Just learnt that there are two types of internal link: relative path and absolute path. This can mess up my blog
ariesta.id/blog/2025/01...
ariesta.id/blog/2025/01...
No trailing slash messes up MkDocs relative link - ariesta.id
ariesta.id
January 6, 2025 at 5:32 PM
Just learnt that there are two types of internal link: relative path and absolute path. This can mess up my blog
ariesta.id/blog/2025/01...
ariesta.id/blog/2025/01...
An attempt to maintain a link blog. First post: ariesta.id/blog/2025/01...
inspired by simonwillison.net/2024/Dec/22/...
inspired by simonwillison.net/2024/Dec/22/...
Flask App with Google Login - ariesta.id
ariesta.id
January 3, 2025 at 5:03 PM
An attempt to maintain a link blog. First post: ariesta.id/blog/2025/01...
inspired by simonwillison.net/2024/Dec/22/...
inspired by simonwillison.net/2024/Dec/22/...
Feels good deploying this with Coolify, Docker, Flask, and Google Login. I'm tempted by the one-click deployment on Coolify that I went through docker configuration hell. Luckily there are Claude, Gemini, Llama, StackOverflow
Yes looks is last priority
github.com/ariesta-id/t...
Yes looks is last priority
github.com/ariesta-id/t...
January 3, 2025 at 2:50 PM
Feels good deploying this with Coolify, Docker, Flask, and Google Login. I'm tempted by the one-click deployment on Coolify that I went through docker configuration hell. Luckily there are Claude, Gemini, Llama, StackOverflow
Yes looks is last priority
github.com/ariesta-id/t...
Yes looks is last priority
github.com/ariesta-id/t...
Why is it so hard for LLM to suggest --no-ff and --orphan for my git needs? Sorry for my broken English but I thought it's a common use case and LLM can easily get what I meant.
January 3, 2025 at 5:31 AM
Why is it so hard for LLM to suggest --no-ff and --orphan for my git needs? Sorry for my broken English but I thought it's a common use case and LLM can easily get what I meant.
Reposted by Alvin Ariesta
Just a reminder that none of the people who make LLMs, no matter how smart, actually know what specific tasks LLMs will be good or bad at. We are barely benchmarking these systems at all on any sorts of tasks.
You should explore in areas of your expertise to try to figure it out for your use cases.
You should explore in areas of your expertise to try to figure it out for your use cases.
December 16, 2024 at 3:14 AM
Just a reminder that none of the people who make LLMs, no matter how smart, actually know what specific tasks LLMs will be good or bad at. We are barely benchmarking these systems at all on any sorts of tasks.
You should explore in areas of your expertise to try to figure it out for your use cases.
You should explore in areas of your expertise to try to figure it out for your use cases.
Reposted by Alvin Ariesta
Sometimes our anthropocentric assumptions about how intelligence "should" work (like using language for reasoning) may be holding AI back. Letting AI reason in its own native "language" in latent space could unlock new capabilities, improving reasoning over Chain of Thought. arxiv.org/pdf/2412.06769
December 10, 2024 at 2:59 PM
Sometimes our anthropocentric assumptions about how intelligence "should" work (like using language for reasoning) may be holding AI back. Letting AI reason in its own native "language" in latent space could unlock new capabilities, improving reasoning over Chain of Thought. arxiv.org/pdf/2412.06769