Hacker News Reader

agnokapatheticon Google deletes net-zero pledge from sustainability website

> Electrical signal attenuation increases with the square of the distance

not true. in standard HV-AC lines, power losses are ~10% per megameter. HVDC gets to 3-5%. So Nevada to Seattle would be at most 20% loss, and in practice 15%, and with HVDC closer to 7%.

https://www.nationalgrid.com/sites/default/files/documents/1...

agnokapatheticon Nvidia DGX Spark

it uses LPDDR5x instead of the datacenter variant’s HBM3e.

agnokapatheticon XMLUI

reminds me of Macromedia MXML (used in Flash, Coldfusion and Flex)!

https://github.com/apache/royale-asjs/blob/develop/examples/...

agnokapatheticon OpenAI’s Windsurf deal is off, and Windsurf’s CEO is going to Google

claude code has a Jetbrains plugin which is delightful!

agnokapatheticon Something weird is happening with LLMs and chess

https://en.wikipedia.org/wiki/OpenAI_Five

https://openai.com/index/gym-retro/

agnokapatheticon Character.ai CEO Noam Shazeer Returns to Google

per The Information -- investors don't get screwed, the board repurchased all shares at $88/share ($2.5B valuation).

agnokapatheticon [dead]

did they try to Google "how to make LSD"? several widely available guides. Tired of LLMs being seen as "risky" for doing the same thing search engines and blogs have been doing for two decades.

agnokapatheticon Data breach at Kaiser Permanente affects 13.4M people

this is just from Share Embeds.

they now disclose these are used at login.

agnokapatheticon HIV in cell culture can be completely eliminated using CRISPR-Cas gene editing [pdf]

mRNA vaccines are about getting mRNA into just a few non-specific cells which then instructs the cell to produce Corona Virus Spike Protein. This gets identified by the immune system which then makes antibodies targeting the spike protein.

So the mRNA only needs to get taken up by a few cells to produce an immune response.

The HIV CRISPR-Cas technique needs to target several orders of magnitude more cells in the body, and is performing a far more complicated procedure than simple transcription like with an mRNA vaccine.

agnokapatheticon Building an early warning system for LLM-aided biological threat creation

They're pretty clear about it though erring on the conservative side.

> While none of the above results were statistically significant, we interpret our results to indicate that access to (research-only) GPT-4 may increase experts’ ability to access information about biological threats, particularly for accuracy and completeness of tasks. This access to research-only GPT-4, along with our larger sample size, different scoring rubric, and different task design (e.g., individuals instead of teams, and significantly shorter duration) may also help explain the difference between our conclusions and those of Mouton et al. 2024, who concluded that LLMs do not increase information access at this time.

agnokapatheticon Ask HN: Is Apple's R1 a Discrete GPU?

it’s a discrete Image Processing Unit

https://en.wikipedia.org/wiki/Image_processor

agnokapatheticon Competitive performance claims and Inference performance on AMD MI300X

batch size 1 -- this is a terrible benchmark that really only shows memory bandwidth only -- LLM inference on Llama2-70b is memory bound up to a batch size of a half dozen or so.

agnokapatheticon The AI Market Is Firming Up Fast

But equities investors value cloud growth for all of these companies, and AI spend is high enough that even if the margin isn't there, the revenue growth will make the street happy.

agnokapatheticon Ask HN: Retrieving multiple images with a single request?

CSS Image Sprites https://www.w3schools.com/Css/css_image_sprites.asp

agnokapatheticon OpenAI is too cheap to beat

> "In some specific tasks, a finetuned 7B llama can work as well as GPT3.5."

"some" is doing a lot of heavy lifting here.

Also: don't discount the labor cost of curating a fine tuning dataset, running a FT training run, even if the hardware is cheap.

agnokapatheticon Slack's Google Drive App can share your private Docs and Drive files

This has been shared with Slack many times by many separate organizations and always closed with WontFix / Working as Expected

agnokapatheticon Elegant Palantir Access Control PowerPoint (2008)

Random aside: The author of this, Bob McGrew, is now the VP of Research at OpenAI.

agnokapatheticon Amazon will invest up to $4B in Anthropic

OpenAI has committed to never training on API data. You no longer need a DPA for this.

https://openai.com/enterprise-privacy

agnokapatheticon Llama 2 on togetherAI is as bad of a privacy nightmare as OpenAI

It's quite easy to opt-out of OpenAI using your data for training. API and ChatGPT Enterprise are private by default.

# OpenAI Products and Privacy

- ChatGPT Free - trains on your data unless you Opt Out

- ChatGPT Plus - trains on your data unless you Opt Out

- ChatGPT Enterprise - does _not_ train on your data

- OpenAI API - does _not_ train on your data

Opt-out link: https://help.openai.com/en/articles/5722486-how-your-data-is...

agnokapatheticon Mullvad on Tailscale: Privately browse the web

A lot of the people making purchasing decisions to acquire products like Tailscale are in security departments and have a very low opinion of Mullvad (VPN of choice for all kinds of abusive/fraud/hacking traffic).

agnokapatheticon ChatGPT Enterprise

This would be a violation of GDPR and CCPA and would expose them to massive litigation liability.

agnokapatheticon ChatGPT Enterprise

There are two different mechanisms here:

Disabling "Chat History / Training" in ChatGpT Settings will disable chat history.

Opting out through the linked form in that FAQ will allow you to keep chat history.

agnokapatheticon ChatGPT Enterprise

Because it's clear as mud from a privacy perspective:

# OpenAI Offerings

- ChatGPT Free - trains on your data unless you Opt Out

- ChatGPT Plus - trains on your data unless you Opt Out

- ChatGPT Enterprise - does _not_ train on your data

- OpenAI API - does _not_ train on your data

# Microsoft Offerings

- GitHub Copilot - trains on your data

- GitHub Copilot for Business - does _not_ train on your data

- Bing Chat - trains on your data

- Bing Chat Enterprise - does _not_ train on your data

- Microsoft 365 Copilot - does _not_ train on your data

- Azure OpenAI Service - does _not_ train on your data

Opt-out link: https://help.openai.com/en/articles/7039943-data-usage-for-c...

agnokapatheticon Beating GPT-4 on HumanEval with a fine-tuned CodeLlama-34B

GPT-4 as served in the API has been getting 85% on HumanEval (compared to 69.5% claimed here)

https://twitter.com/amanrsanger/status/1635751764577361921 https://github.com/getcursor/eval

agnokapatheticon Llama2 on Replicate faster than ChatGPT?

The response time you get from a Pay-as-you-Go account on OpenAI is not anything close to what you get with an Enterprise Agreement. If your usage is large enough, talk to your account rep.

agnokapatheticon A room-temperature superconductor? New developments

That 40kJ/L is for the entire system including the cyronics. Get rid of the cyronics and the system gets far far more energy dense.

agnokapatheticon Judeo-Christian roots will ensure US military AI is used ethically, general says

No. The Notion was not very much the same. Medieval views of the Torah / Old Testament were very much informed by anti-semitism.

Medieval crusaders would have balked at the idea that Jewish and Christian values could have any form of harmony.

agnokapatheticon Dutch E-Bike startup VanMoof unable to pay the bills

They're essentially just doing light final assembly/packaging on Bafang ODM subassemblies.