mkolodny

Google has been publishing frontier research since it was an early-stage startup (https://research.google/pubs/the-anatomy-of-a-large-scale-hy...).

Good idea! It could be fun to try creating an RSA-esque visual explanation video instead of using canvas selection (https://youtu.be/zDZFcDGpL4U).

Also reminds me of 3D Ultra Minigolf Deluxe (https://youtu.be/mr9GbmBROfo). That game was kind of Disney World meets minigolf with how imaginative the levels were.

This is super helpful :) Curious about Grok as well!

Even if it’s not perfect, I’m happy to see there’s a focus on AI Security. NIST has been a reliable producer of quality international standards for cybersecurity. Hopefully this action plan will lead to similarly high quality recommendations for AI Security.

Humans have limited ability to self-introspect, too. Even if we understood exactly how our brains work, answering “why?” we do things might still be very difficult and complex.

Biases can lead to opinions, goals, and aspirations. For example, if you only read about the bad things Israelis or Palestinians have done, you might form an opinion that one of those groups is bad. Your answers to questions about the subject would reflect that opinion. Of course, less, biased information means you’d be less intelligent and give incorrect answers at times. The bias would likely lower your general intelligence - affecting your answers to seemingly unrelated but distantly connected questions. I’d expect that the same is true of LLMs.

Did you read the comment I’m replying to?

> They have LITERAL human editors choosing which stories to put on the front page, just like the NYT, and they should be held liable for the content on their platforms just like legacy media is.

Do you think a Facebook feed is just like the New York Times?

Are you saying you think human editors handpick what goes on billions of people’s Facebook feeds?

“Got caught” is a misleading way to present what happened.

According to the article, Meta publicly stated, right below the benchmark comparison, that the version of Llama on LMArena was the experimental chat version:

> According to Meta’s own materials, it deployed an “experimental chat version” of Maverick to LMArena that was specifically “optimized for conversationality”

The AI benchmark in question, LMArena, compares Llama 4 experimental to closed models like ChatGPT 4o latest, and Llama performs better (https://lmarena.ai/?leaderboard).

Is that what DOGE is doing? Helping decide the SEC’s policies?

> The SEC has its budget set by Congress, but the actual funding comes from transaction fees imposed on the financial sector, meaning its operations ultimately cost the taxpayer nothing, according to the agency.

How do you make something that costs nothing more efficient?

The same government where the head of the department of efficiency is also the owner of Anthropic’s and OpenAI’s competitor, xAI?

> security reasons later

What about security reasons now? The federal government includes the military. Giving DOGE “God mode” on the federal government is a national security risk right now.

A vague “stuff is happening behind closed doors” isn’t enough of a reason to build AI weapons. If you shared a specific weapon that could only be countered with AI weapons, that might make me feel differently. But right now I can’t imagine a reason we’d need or want robots to decide who to kill.

When people talk about AI being dangerous, or possibly bringing about the end of the world, I usually disagree. But AI weapons are obviously dangerous, and could easily get out of control. Their whole point is that they are out of control.

The issue isn’t that AI weapons are “evil”. It’s that value alignment isn’t a solved problem, and AI weapons could kill people we wouldn’t want them to kill.

PRELIMINARY EARTHQUAKE PARAMETERS ---------------------------------

* The following parameters are based on a rapid preliminary assessment of the earthquake and changes may occur.

* Magnitude 7.3 * Origin Time 0944 AKST Dec 05 2024 1044 PST Dec 05 2024 1844 UTC Dec 05 2024 * Coordinates 40.3 North 124.7 West * Depth 8 miles * Location 45 miles SW of Eureka, California 215 miles NW of San Francisco, California

One vendor on VirusTotal flagged the domain currently linked from their Wikipedia article as malicious.

On VirusTotal, 5 different vendors flag Z-Library as malicious. Are they just flagging the site because of IP issues, or is the site full of malware?

The Internet was largely developed by the U.S. Defense Advanced Research Projects Agency (DARPA) [0]. Tools like the Internet and AI can be used for many things, good and bad.

[0] https://www.internetsociety.org/internet/history-internet/br...

“You may be surprised to learn that of the trio of long-awaited coronavirus vaccines, the most promising, Moderna’s mRNA-1273, which reported a 94.5 percent efficacy rate on November 16, had been designed by January 13. This was just two days after the genetic sequence had been made public”

https://nymag.com/intelligencer/2020/12/moderna-covid-19-vac...

I'm surprised people are still making that argument even after the pandemic showed us the risk is nowhere near worth the reward. Regardless you need to create a vaccine for the discovered virus (which can take less than a week, as was the case with covid). And then you still need to go through months of human trials.

I was hoping we were done risking starting pandemics by purposefully creating new deadly viruses.

I disagree. The file I linked to, model.py, contains the Llama 3 model itself.

You can use that model with open data to train it from scratch yourself. Or you can load Meta’s open weights and have a working LLM.

Llama’s code is open source: https://github.com/meta-llama/llama3/blob/main/llama/model.p...

Given that blind people can learn to speak, audio alone must be enough to learn language. And given that deaf people can learn sign language, video alone must also be enough to learn language. That’s assuming that touch and emotion aren’t crucial to language learning.

That's just the default. You can set max_seq_len to 8k. From the readme [0]:

> All models support sequence length up to 8192 tokens, but we pre-allocate the cache according to max_seq_len and max_batch_size values. So set those according to your hardware.

[0] https://github.com/meta-llama/llama3/tree/14aab0428d3ec3a959...

Ex-Uber fullstack and machine learning engineer with over 10 years of experience.

Location: New York City, USA

Remote: No

Willing to Relocate: No

Technologies: Python, Django, TypeScript/JS, React, Go, Swift, C, HTML, CSS, SQL, PyTorch, TensorFlow

Résumé/CV: https://www.linkedin.com/in/michaelkolodny/

Email: michaelckolodny [at] gmail [dot] com

If it's true, condolences and love to his family.