kartoolOz

It's very hyper-parameter dependent, and in my testing didn't provide comparable performance to maxsim.

"कर्मण्येवाधिकारस्ते मा फलेषु कदाचन | मा कर्मफलहेतुर्भूर्मा ते सङ्गोऽस्त्वकर्मणि" - Bhagvad gita, chapter 2, verse 47.

You have a right to perform your prescribed duties, but you are not entitled to the fruits of your actions. Never consider yourself to be the cause of the results of your activities, nor be attached to inaction.

WebArena does this really well, called the "accessibility_tree" https://github.com/web-arena-x/webarena/blob/main/browser_en...

The Info extraction and Question Answering metrics are far worse than transformers though.

They also say that in the blog "However, both Based and Mamba still underperform the strongest Transformer baseline, sometimes by large margins. This is consistent with our “no free lunch” observation above"

Alphacode 2 technical paper claims to solve 43% of problems (77 problems from 12 codeforce competitions) performing 85%ile on all human participants.

Caveat is deep in the technical paper,

1) generates 1m candidates from N different prior models (fine tuned on previous codeforces)

2) throws away 95%ile candidates (doesn't fit the test case + no compile)

3) groups semantically similar candidates

4) scores candidates from each group (Based on another scoring model, probably latency + descriptions etc)

5) Picks top 10

Makes 10 submissions and finally gets the score ..

sure this is how humans solve problems ... totally awed by AGI /s

Technical report: https://storage.googleapis.com/deepmind-media/gemini/gemini_... Nano-2 is 3.25b, and as per figure 3, nano-2 is roughly 0.6-0.8 as good as pro, and ultra is 1.05-1.3 as good as pro.

Roughly that should put gemini ultra in the sub 100b range?

Hi, Thanks for Open sourcing the code! I was trying to reuse the code especially the dynamic quantization per channel (int8 on gpu) but couldn't get it to work, i also checked out torchao package but it looks like it has dependency on the nightly channel and SAM's dynamic implementation with triton has other issues, is there any clean implementation of int8 dynamic post-training quantization that you can point too ?

Would see great improvments in retrieval accuracy by finetuning e5-base-v2 or the newer leaders on mteb benchmark.

Vespa.ai is pretty crazy too, a bit unkown we run a huge vespa cluster serving 1k+ queries with <100ms latency ...

Hindu - is someone who practices Hinduism and identifies as such, its not a language, that's Hindi.

It's also rudimentary to think in binary that you can either be a programmer or a manager. Progressing in my career i am finding this duality to be incredibly constricting, I excel best at programming, machine learning & working with business, product teams in finding areas of highest impact and then executing with smallish teams but i am increasingly finding that this doesn't fit into most corporate structures and i find resistance in landing the right role.

ML engineer with 10+ years experience, took voluntary break in Sep, started interviewing in Nov, interviewed at 6 companies to land 2 offers.

Most companies have been having hacker-rank pre-screen coding, Hiring Manager Technical screen, System Design, Product round, Behavioral round, ML specific round.

While it's not been hard to land interviews, the process is taking a long time and it looks like people are looking for previous experience in the specific stack they work on. Offers are also lower from a year ago.

Any idea how does it perform compared too Cassandra, Scylla other cloud offering wr.t maintainence, scaling, throughput, cost effectiveness ?

These are relatively easy to build and can be used for a variety of tasks like Entity Resolution, https://news.ycombinator.com/item?id=32825679

Even self hosted milvus supports this.

It's supports all CRUD operations, used pinecone as the vector storage and it's highly available and can be scaled to additional pods, existing solution was using zentity.

I recently had to solve the Entity Resolution Problem at my workplace, and here was how i went about it.

Problem Statement:

   - Find the right entity for a query among ~50m entities.

   - Queries could be few of the entity attributes (If entities have n attributes, query can in 1..n)

   - Queries can mention attributes in various ways (Partial information, Typing errors, abbreviation, Extra information etc)

Existing Solution:

    - Elastic search based match, using complicated heuristics overfit on a small training set. Gets worse over time as number of entities increase, the top-20 search retrieval accuracy was around ~40% on current number of entities.

Implemented Solution:

     - Embedding search using a Sentence embedding model (pretrained Deberta finetuned for current problem) trained via Contrastive Learning, where positive pairs are generated using augmentations for each attribute which best mock the queries (after going through many user queries)

     - Top-20 accuracy was around 98%, filtering out right entity was through hueristics and other business logic with proper confidence measure (hyper-param tuned on val set), after final pipeline we could get the high confidence top-1 accuracy to around 99.995% (precision) and 86% (recall).

We ended up going with pinecone for the embedding search, and the search latency was around ~100ms (top 50 among ~50m embeddings)

Negotiating the title and then researching salary bands for that title probably will give you more insight into what to expect, once you have a better understanding of their bands, asking a number in top 10% of the particular band should be doable. If the number doesnt impress you, ask for a higher title.

Sidenote, been using pinecone for semantic search recently, and it's been a joyful experience.

Single X-ray session costs 4$ in my state in India

thanks for the reply, i used to be an sde2 at amazon previously, currently working at a top 50 by market size indian company.

AMD 5900x, yup running a single process

Tried fastwsgi + falcon cythonized.

  Running 1m test @ http://localhost:5000
    8 threads and 100 connections
    Thread Stats   Avg      Stdev     Max   +/- Stdev
      Latency     1.41ms  109.36us   3.85ms   76.27%
      Req/Sec     8.53k   625.45    25.69k    79.77%
    Latency Distribution
      50%    1.43ms
      75%    1.48ms
      90%    1.53ms
      99%    1.62ms
    4072168 requests in 1.00m, 454.37MB read
  Requests/sec:  67816.66
  Transfer/sec:      7.57MB

Pure fastwsgi on same machine

  Running 1m test @ http://localhost:5000
    8 threads and 100 connections
    Thread Stats   Avg      Stdev     Max   +/- Stdev
      Latency   645.52us   30.50us   1.33ms   82.06%
      Req/Sec    18.67k   795.08    36.56k    86.86%
    Latency Distribution
      50%  648.00us
      75%  658.00us
      90%  668.00us
      99%  736.00us
    8919901 requests in 1.00m, 867.68MB read
  Requests/sec: 148419.82
  Transfer/sec:     14.44MB

Damn nice.

How big tech got so big, passive funds.

Employees are replacable, founders would always maximize money first.

Author of the symspell port for python using c++/pybind11 https://github.com/viig99/SymSpellCppPy

Modern spell checkers are fairly interesting beasts. Do check out, use and let me know if i can help with any issues / features.

Posts like this make me want to sell my soul to the devil O.O

I feel the whole woke culture has led to lesser mentorship opportunities for women, most senior guys I know are too afraid to mentor in this climate.