131 karma · joined June 23, 2016
What is great is that you can define DSPy signature of the type “question, data -> answer” where “data” is a pandas dataframe, then DSPy prompts the llm to answer the question using the data and python code. Extremely powerful.
I was in your situation in 2023 and decided to quit without a job lined up. I rested for 2 months and then, out of nowhere, I had energy and motivation to start exercising again and start working on my mental health (with the help of a therapist/professional coach). I would recommend to have one single expectation if you decide to quit: rest.
All in all, quitting the toxic workplace and taking time off was the best thing I’ve ever done.
It took me 7 months to get an offer, 9 to actually start working.
Be prepared to be out of work for ~1 year.
Because data must be on-prem, banks are stuck in legacy infra paradigms. The whole org suffers, innovation is stiffled, yada yada…
An on-prem cloud product (hardware+software) is a game changer for these companies, IMO.
My question to oxide: how easy is to integrate external hardware into the cloud? For example: bunch of GPUs or a bunch of next-gen hardware like SambaNova.
I’d be really good to allow more than two models and change dynamically based on multiple constraints like latency, reasoning complexity, costs, etc.
The answer is here. This is a cost-saving tool.
All companies and their moms want to be in the GenAI game but have strict budgets. Tools like this help to keep GenAI projects within budget.
A few questions I jotted down while watching the video on Struct's landing page:
1. the concept of channels seems to be important on Struct as channels are the starting point of threads/feeds. Could you clarify the concept of channels on Struct? Is it just a concept to group users? Can you also chat on channels?
2. Conceptually how do you handle the fact that only the threads on the realtime feed are visible to the user? Maybe there's a low-signal high-activity thread that takes space and hides the high-signal low-activity thread which results in users missing important information or reminders.
3. Tags are crucial for filtering threads, is there a way to "police" the tags? Using tags usually grow into a mess of similar-but-not-the-same collection of text. Think of JIRA tags.
4. How to handle threads created independently by different users but discussing the same topic?
5. Not a question, but I'd be interested in knowing more about private conversations between two parties. It's mentioned only briefly in the video.
Hopefully these questions don't come out as overly critical. The tool definitely has potential.
With that out of the way, basic linear algebra operations that require sophisticated algorithms and are not "embarrassingly parallel": matrix multiplication, matrix inversion, matrix decomposition (SVM, QR, etc). Some of these algos fall into BLAS and others in LAPACK.
It's the exact opposite, most numerical linear algebra is _not_ embarrassingly parallel and requires quite an effort to code properly.
That is why BLAS/LAPACK is popular and there are few competing implementations.
This man needs to learn to edit himself.
The article also mentions that the kid didn't have the symptoms or physical manifestations of spina bifida, complicating the diagnosis.
This case reads as a genuinely difficult case to diagnose.
Top talent from, say, Vietnam already speaks good English and many would jump to the opportunity to work in Japan under these conditions.
What kind of problems do mean? Environmental? Or health-related problems?
1. SEON: https://seon.io/
2.IPQS: https://www.ipqualityscore.com/
The author explains that happiness is subjective:
> The behavioral science literature often refers to happiness as subjective well-being because the meaning of happiness varies in different contexts
...and then proceeds to delineate how psychology defines happiness:
1. a person’s own assessment of their satisfaction with life;
2. how much positive emotion [...] they experience;
3. and how little negative emotion [...] they experience
So yeah, self-rating is an important factor to measuring happiness.
I see advertising radically different.
People exhibit an spectrum of interest in products in the market, from “zero interest in buying” to “shut up and take my money”. Advertising works by convincing people close to the “shut up and take my money” part of the spectrum to actually buy.
Disclaimer: I’m not a marketeer.
> The separate guilty pleas entered by the hackers demonstrate that after Sullivan assisted in covering up the nature of the hack of Uber, the hackers were able to commit an additional intrusion at another corporate entity—Lynda.com—and attempt to ransom that data as well.
While data.table is faster than dplyr, data manipulations with data.table are difficult to read/understand/maintain.
dplyr also grew into a full-fledge list of libraries to work on data-related projects (the tidyverse). These libraries are _very_ well thought out and enables productivity with minimal learning curve [anecdotal]
It would be a killer product because most users are familiar with spreadsheets already. Many users end-up copying data from dashboards into spreadsheets (gsheet, excel, etc) so why not skipping the intermediaries and go straight to delivering a hybrid of dashboards and spreadsheets.
Lots of potential.
It is still undetermined if having 12 highly-skilled professionals in the experiment is enough to have a conclusive experiment.
Also, this subject is so difficult to get right that the authors of the article themselves hedged by saying that experiment "does not support watertight conclusions".
[Anecdotal] One example of the difference between MP3 and lossless: the "image" [1] on 256kbps MP3s is worse compared to the the original uncompressed, lossless, versions (but the listening room must be appropriately prepared to reproduce a good image).
This is a highly subjective topic. IMO we'll never reach full agreement. Personally, I listen MP3 while on-the-go and lossless music at home.
Important to keep in mind the "size" of the experiment. Two interesting quotes from the article in c't magazine:
> twelve participants would be asked to come to Hanover.
> It's true that the data we collected does not support watertight conclusions, but they do provide interesting insights.
There are at least four issues with this advice (wrt doing data analysis):
1. How do you link your PCA components to the original data? Let's say you are tasked to find the main drivers of sales on a given city. You run PCA on the data and find two main components on the dataset. What do you do next? How do you make this information actionable?
2. How do you treat categorical variables? There are PCA methods for dealing with categorical variables but by the time you apply these methods plus the issues in 1) your data has lost all actionable meaning.
3. PCA is _very_ difficult to explain to business stakeholders. The more difficulty business stakeholders have to understand the analysis, the less they will use it.
4. Data-driven business stakeholders will favour clarity and simplicity over sophistication (somewhat linked to 3)
This is not good advice.
I’d say find experts that you trust and be aware that these experts may commit mistakes as any other human.
In short: find the experts that make the least mistakes.
As good as LibreOffice may be (I personally think it is _not_ better than Google docs) and as much as I like the romantic idea of open source office suite, I don't believe this will be successful/sustainable in the long run. The bulk of office suite users are not interested on using alternatives. These users want office running on their machine, period. I'd even venture to say that most users would rather have apps installed instead of Microsoft Office Online.