HNHacker News
TopNewBestAskShowJobs

algo_trader

565 karma · joined January 15, 2018

algo_trader2017@protonmail.com

Security, algorithms and algo trading since 19XX..

submissionscomments
algo_trader··on Sky lapse in two tone
Off topic: are there datasets for solar incidence over the oceans ?

There are plenty of accurate maps/widgets for ground based PV, but ocean data seems much rarer

algo_trader··on ECJ finds IAB Europe responsible for TCF consent spam popups across the Internet
> Massive win for Google, Apple, Facebook.

Yes, that is true and under appreciated

> Really hard to see a future for third party ad networks

For now, what are biggest programmatic exchanges still going? I have been out of the loop for a while

algo_trader··on Mamba Explained: The State Space Model Taking On Transformers
> at least provide evidence that scaling up no longer works for transformers.

Isnt the Mixture-of-Experts trend (GPT4 is MoE?) kinda of a proof ?

algo_trader··on Mamba Explained: The State Space Model Taking On Transformers
> Just that it isn't the immediate transformer killer.

What is the best/stable-ish linear alternative for transformer right now? Especially for text generation and summarization.

We have domain specific ways of over sampling and search, so we much prefer less expensive models.

algo_trader··on Building a deep learning rig
together.ai is really good but there is a price mismatch for small models (a 1BN model is not x10 cheaper than 10BN models)

This is obviously because their are forced to use high memory cards.

Are there ideal cards for low memory (1-2BN) models? So higher flops/$ on crippled memory

algo_trader··on Building a deep learning rig
> built an AI rig and went with 2x3090s,

Is there a goto card for low memory (1-2BN) models?

Something with much better flops/$ but purposely crippled with low memory.

algo_trader··on Owners report rust forming on Tesla Cybertruck
What is the cheaper alternative for 316 in marine applications?

Specially long term floating objects such as buoys, with minimal physical contacts, and a painting/maintenance schedule is possible

algo_trader··on Lessons Learned from Payments Startups
Nice write up, and i hope you dont regret the journey

I am surprised you didnt mention fraud - is it because micro-payments are not "fraud worthy" or because it is a manageable problem ?

algo_trader··on ALOHA robot learns from humans to cook, clean, do laundry
1-human tele-operating 10-robots seems like an obvious solution Is any one working on this?

Certainly for most commercial/industrial/agriculture/hotels/mining settings (in fact, everything except a private apartment butler thingy..?)

algo_trader··on Slow Electricity: The Return of DC Power? (2016)
If you are building a custom mini grid from scratch, how do you get from batteries/panels to distribution-level 11kv DC without going through AC ?
algo_trader··on Paving the way to efficient architectures: StripedHyena-7B
> H 7B is always faster than optimized Transformers > (>10%, >20% and >50% end-to-end faster than FlashAttention

50% improvement on very large sequences is great. But it is not yet transformative (pun intended..)

Presumably FlashAttention already does lots of the less-than-quadratic improvement

algo_trader··on Fine-tuning Mistral 7B on Magic the Gathering Draft
> https://www.youtube.com/watch?v=Xq4T44EvPvo

Why would he need to write a game simulator from scratch?

Surely there are OSS versions which are "good enough" ?

(disclosure: not a MtG expert)

algo_trader··on Exponentially faster language modelling
Axelera [1] or Halio [2] give you 100-200tflop for ~$200.

8-bit ops, inference only, low memory embedded, excluding the host, implied utilization from FPS specs is ~20%

But the trend is there.

There are also newer ADAS/AV units from China which claim 1000tflops and cant really cost more than $1000/$2000 per car.

These are all tiled designed (see also dojo/tesla) heavily over-weighed on flops vs memory

[1] https://www.axelera.ai/

[2] https://hailo.ai/

algo_trader··on Exponentially faster language modelling
Infact, as stated in the paper, this is bad news

> We therefore leave the attention layers untouched

Meaning, presumably, that the GPU memory remains the bottleneck

Flops really are quite cheap by now, e.g. vision inference chip ~$2/teraflop/s !!

algo_trader··on XRDP with H.264 codec brings immense performance boost
Off topic, and apologies

But is there a recommended client for chromebooks?

algo_trader··on Citibank analyst dismissed for lying about meals expenses claim under €100 limit
> The senior manager questioned his answer, saying: > “The receipt appears to have two sandwiches, two coffees, and another drink . . > Are you advising that this was all consumed by you?”

The entire article is hilarious

algo_trader··on Daugher of Mellanox founder killed in Hamas attack
> Danielle Waldman, who was born in Palo Alto, California, was attending the music festival with her boyfriend of six years

> ".. From the shells that we have found, there were at least three guns that were shooting at the car"

A horrible barbaric tragedy

Rest in peace

algo_trader··on How Netflix uses Java
> Virtual threads are incredibly impressive work,

Do you have an (un)informed opinion on minimum task sizes for the green threads?

My interest is re-factoring java code to reduce total wall clock time, on large compute with plenty of memory/cache.

algo_trader··on The casino in your pocket
Thank u for your blog. Very thoughtful with helpful references.

Any books u recommend on gamifications?

algo_trader··on Q-Transformer: Scalable Reinforcement Learning via Autoregressive Q-Functions
I havent yet fully grokked RWKV..

Just how much compute/memory are we saving here?

My understanding is that a 1BN transformer is about 2BN flops/inference, so about 1TFLOP for a 500 sequence of inferences (and also about several GB of memory)

What would be the equivalent RWKV (let ignore the inevitable loss penalty which could be significant..)

algo_trader··on Exllamav2: Inference library for running LLMs locally on consumer-class GPUs
Thanks for the writeup

Our special case is a single model, high throughput and latency unsensitive.

Can we use multiple lower-memory GPUs, split the the model "horizontally", and pipeline each batch/inference across the GPUs?

Is this a common scenarios handled by the engines you mentioned ?

algo_trader··on DC Grids
> but this is actually my day job.

You are probably more knowledgable than me. I am just here for the cake.

However, AFAIK the most fault prone components in PV are inverters and (where applicable) transformers.

> 5-10x more than what was mentioned.

This was my mistake. I started with 5c/W and missed a decimal place.

None of this addresses the key issue - you mentioned 10% loss for a simple DC converter, and the next commenter [1] mentioned the newer generation of far superior alternatives.

[1] https://news.ycombinator.com/item?id=37260441

algo_trader··on DC Grids
Yes. The original commenter is a bit behind the times. There is also reliability advantages.

> new(er) GaN(or SiC) DC/DC buck conversion

Can you share guide lines about costs?

For comparison, utility scale AC inverters for PV are ~$5-$10/kW

algo_trader··on BHP says battery electric cheaper than hydrogen as it dumps diesel
Please see my sibling comment.

Do we really use 2000A cables running around a work site? Are they buried ?

algo_trader··on BHP says battery electric cheaper than hydrogen as it dumps diesel
> fat ass HV cables running to them

Do you a more exact specs?

Maybe MV 10kV-40kV ?

And how thick are these cables?

Do they carry 2000A snaking around the site?

algo_trader··on China Launches First 700 TEU Electric Containership for Yangtze Service
> The problem is scale.

Well yeah.

But batteries are the still most expensive component - much more than transmission/charging upgrades.

So if u do hot-swappable batteries, this component is only cycled once every 2-3 days, while you wait for the next ship.

Power will cost you 5-10c/kwh, and 10c/kwh for battery capex amortization ($100/kwh LFP with 100 cycles/year).

Sigh

algo_trader··on Waymo puts the brakes on self-driving trucks program
Which do you think has a stronger moat?

I have a toy model for e-trucking + autonomy + middle mile. But its a long hard slog to roll out (ha! pun!)

algo_trader··on Superconductor news: What’s claimed, and how strong the evidence seems to be
> (as well as an obvious instant Nobel prize).

South Korea has been waiting for a (real, STEM) Nobel prize for decades.

These guys will be worshipped as gods if they deliver

algo_trader··on Crucial system of ocean currents is collapsing; ‘affects everyone on the planet’
> if we reduce emmissions to zero this year, it still results in +1.5 C increase by the end of the century

this is non-sense according to the ipcc models

for 1.5C, we still have a budget of 200BN tonCO2e

for 2C, we have about 1TN tonCO2e

for 2.5C, 2TN ton, is within current trajectory of progress

we currently emmit about 50BN tons/annum

All of the above are bad, but degrees (ha! pun!) matter

> China's overfishing

Bashing the developing world is not helpful. The US account for a massive chunk of historical emissions and excess consumption.

algo_trader··on Hotel Owners Start to Write Off San Francisco as Business Nosedives
Did you come across any data on (up to date) construction costs ?
← PreviousPage 5 of 13Next →