yarri

>> To be clear, pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this

> if slowing this down were possible

Why is embedded alignment evaluation not possible?

I agree with Dario, this did work globally in banking and did encourage a race to the top. Didn’t prevent the GFC but also we did survive the GFC.

Pallas is the Triton equivalent in JAX land. There are some old auto tuning prototypes if you search for Pallas, like this https://github.com/jax-ml/jax-triton/pull/108

Many of us grew up in the PLD era, k-maps, etc. Woz pushed early HW to the limit, with SW APIs that delivered real value. Woz made astute design trade-offs based on full stack knowledge that his peers lacked. The world’s moved on to the GPU (low precision, accelerated parallel compute?) era, but the Woz view point still holds. You can see it in the AI kernel optimizations, or rematerialization methods to push GPU HW to the new limits, and trade-offs need to be made. GPU HW for 4-bit QAT or even 2-bit will dramatically affect the SW (AI) of this era. What trade-offs do you make?

I saw Woz on Northbound 280 “driving” his cherry red Model S, using FSD. He was looking down at the screen the whole time I watched him. Swear he had ssh’d into it.

[edit - Gabe responded]. See this Cloud Run spending cap recommendation [0] to disable billing, which potentially irreversibly deletes resources but does cap spend!

[0] https://cloud.google.com/billing/docs/how-to/disable-billing...

Not sure what “official” means but would direct you to the GCP MaxText [0] framework which is not what this GDM paper is referring to but rather this repo contains various attention implementations in MaxText/layers/attentions.py

[0] https://github.com/AI-Hypercomputer/maxtext

I assume the Gemini results are JAX/PAX-ML/Pallas improvements for TPUs so would look there for recent PRs

Background on the Tokyo government’s digital twin program, including sourcing and maintenance efforts

https://github.com/tokyo-digitaltwin/roadmap_v1.0/blob/main/...

Please don’t do this, Zach. We need to encourage more investment in the overall EDA market not less. Garry’s pitch is meant for the dreamers, we should all be supportive. It’s a big boat.

Would appreciate the collective energy being spent instead towards adding to amor refining Garry’s request.

This is a datacenter chip. HVAC requirements are more interesting IMO, they seem to be targeting air cooled air edge deployments with that card. They’ll probably wind up with a baseboard design similar to the early v4i TPUs.

https://ieeexplore.ieee.org/document/9499913

Details from their technical memo at https://www.etched.com/announcing-etched

## How can we fit so much more compute on the silicon?

The NVIDIA H200 has 989 TFLOPS of FP16/BF16 compute without sparsity. This is state-of-the-art (more than even Google’s new Trillium chip), and the GB200 launching in 2025 has only 25% more compute (1,250 TFLOPS per die).

Since the vast majority of a GPU’s area is devoted to programmability, specializing on transformers lets you fit far more compute. You can prove this to yourself from first principles:

It takes 10,000 transistors to build a single FP16/BF16/FP8 multiply-add circuit, the building block for all matrix math. The H100 SXM has 528 tensor cores, and each has $4 \times 8 \times 16$ FMA circuits. Multiplying tells us the H100 has 2.7 billion transistors dedicated to tensor cores.

*But an H100 has 80 billion transistors! This means only 3.3% of the transistors on an H100 GPU are used for matrix multiplication!*

This is a deliberate design decision by NVIDIA and other flexible AI chips. If you want to support all kinds of models (CNNs, LSTMs, SSMs, and others), you can’t do much better than this.

By only running transformers, we can fit way more more FLOPS on our chip, without resorting to lower precisions or sparsity.

## Isn’t memory bandwidth the bottleneck on inference?

For modern models like Llama-3, no!

Likely this one… https://en.wikipedia.org/wiki/Scrambling_(military)

- Zuck will spin out FB & Instagram and merge with Twitter : TwiGramFace

Take small bets, explain the value you are attempting to deliver and basically learn how to sell. Especially to skip levels. Be willing to fail.

We all still do a lot of lab work

The name is cool. That logo, not so much.

The name choice is so close...

The rise of punitive solutions is real. I was involved with discussions with local municipalities placing (private) local schools under restrictions for not providing sufficient carpool coverage -- levy fines based on percent of families carpooling.

Would the inverse of these punitive solutions, ie., encouraging carpool / ridesharing, not also work? It always amazes me how relatively unutilized the HOV lanes are.

Depends on the industry. The auto insurance industry had facilitated this, but was then reprimanded for using generic parts to repair damaged cars. There was some irony in that US parts manufacturers claimed 3rd parties were importing "foreign" parts, but many of the US manufactures also subcontracted overseas. Thorny issue.

Civic engagement in Japanese agriculture was a challenge for rural communities when I was living in Japan, this village's activity seems similar to 4H clubs in the US.

Thanks, Kevin, for the time & the detailed feedback.

> the problems with your flow are clearly posting calls to actions on the page Sigh, agreed. I'd like to move away from a wizard-based flow and try either a) using a calendar to drag & drop the schedule, or b) create a WYSIWYG "signup page builder" flow.

>I’d just ask for where is the ultimate destination, maybe a date and a list of emails.

So try to get quicker to the step where a signup page is shared with users, right?

> If I have time, I’ll play around with this on my phone later. Thanks!

Hi, a suburban mother of school aged kids asked me to help her solve the problem she has coordinating with other parents to arrange rides to shuttle their kids to school & sporting events. She wanted a simple scheduling service, with a panic button — a way for her to let others know if she was running late.

Ironically, she works at Redpoint VC, but had trouble attracting developers to her idea. I built the MVP, conducting UserTesting.com focus groups to gather feedback from a busy, educated “soccer mom” demographic and iterated on a “wizard based” signup flow. Users reported it being better than the current email / spreadsheet / shared calendar approach they use now, some interest in paying for the service on a subscription basis. Key request was a native mobile app version, and, confusingly, a Craig's List like rideshare social matching service.

MVP is built on Parse.com / Backbone / Bootstrap, waiting for the right time to build out an iOS app; likely the start of a new school year.

Obviously the growth opportunities are in service-based, market-making transportation providers like Uber, BlaBlaCar/Carpooling.com, whereas this is more of a scheduling app with a long history of similar but failed approaches (as a quick search of ‘carpool’ on HN will show!)

Still, I like the feeling that I'm helping make other people's lives easier.

Feedback & ideas for good mobile app design most welcome!

@sama actually addressed HW on his recent AMA here [0] but as with software startups, you have to broaden your view of a hardware startup -- if you include ventures focused on fundamental semiconductor technology, both analog and digital; tools for design, prototyping and verification; semiconductor intellectual property as well as end products, there are quite a few potential $10B opportunities[1]. I believe @sama is considering the full range of investment options. Not all investments in this segment are capital intensive.

[0] https://news.ycombinator.com/item?id=9238839 [1] http://blog.semiconductors.org/blog/what-end-use-application...

Hesitate to post an AskHN about this, but which company has the 66 year old (co-)founder? [0] Glad to see this diversity.

[0] http://venturebeat.com/wp-content/uploads/2015/03/xlarge_YC-...

Nice job & congratulations on shipping! Some brief feedback before I start field testing :-) this...

- Would love the ability to save a collection of favorites, maybe add folders or tagging?

- I often get asked to create list of recommended craft beers and so would like the ability to share these lists easily

- Your data will be sparse to start with, but I'd be interested in knowing how many other users put beers into their favorites (I find the BeerAdvocate listings a bit tedious to wade through...)

- Not sure you want to, but the professional brewers I speak to are interested in a lower-cost alternative to Untappd [0] and you might be able to build a business here?

- Maybe this is an East Coast US thing, but there's a growing trend to pair beers with cheese [1]; maybe that's too specific a request but allowing used to add notes to the beers (meta data beyond BeerAdvocate) might be useful

Good luck!

[0] https://untappd.com/business [1] http://www.huffingtonpost.com/2014/02/25/beer-cheese-pairing...

Integration with accounting software, like Wave[0]

[0] http://docs.waveapps.io/index.html

I've been thinking of the same. Trying to get my young daughter involved in the process seems the challenging part; weighed against the value of spending time with her on other projects unrelated to gaming.

I want to change the industry, and make a dent in the misogyny and cynicism, but I think the male bias shown here of creating a platformer with a girl hero (antihero?) is speculative. Boys play platformers. My daughter is fine with puzzle games. And dragons. Really cute dragons.

For what it's worth, the game design I worked on put the girl hero as the object of affection of boy enemies and monster boss characters, team up with other girls the thwart the advances of male characters. Misandristic, basically. Which doesn't seem like the best way to make a dent!

If they need the whole weekend to find the right tone to explain this acquisition, yes. Looking forward to seeing how Google handles the privacy concerns on this...

Ummm... I think this just crashed under the HN load:

Warning: mkdir(): No space left on device in /home/tam/Dropbox/Webserver/allthefavicons/htdocs/classes/allthefavicons.class.php on line 19 Warning: mkdir(): No such file or directory in /home/tam/Dropbox/Webserver/allthefavicons/htdocs/classes/allthefavicons.class.php on line 20

Nice work! What are the min hardware requirements for iOS & Android devices?