I deleted my claude account the day the samples started preaching from the pope’s address and claiming 50% of the jobs will be gone while shamelessly stealing entire corpus of human data without attribution while preaching what you can and cannot do and maliciously degrading model quality.
I hope they either never make it to IPO or crash violently.
I have never yet had a set of words that I have grown to hate so much - “taste” and “moat” being at the top of them.
It is almost always coming from people who have or know about neither.
(Agree with the article’s general sentiment - but just wanted to make this tangential comment)
There are alternative ways to gather telemetry data about your usage, then literally polluting the commit message / PR description of the author. Why even consider doing that in the first place?
Of course - finance is the best domain to depkiy a stochastic parrot which hallucinates and forgets stuff frequently and doesn’t follow your instructions - even with SOTA models. One where you need absolute accuracy and auditabikity.
Sheep mentality hard at work at companies. Just because Google does it (processes, technologies, systems etc), lets also adopt it without thinking whether its relevant in our context and use-cases.
I bet the same devs from these firms who are asking to traverse a minimum spanning tree would fumble at even the slightest variation of the problem appearing in daily life.
You prove the point that these are just token generation machines whose output is psuedo-intelligent. It’s probably not there yet to be blindly trusted.
This. I use the same workflow. Also I am too lazy write and maintain notes - so I just use joplin for tags / metadata and typora (wysiwyg editor for md). Thinking and taking notes over typing on Typora is a godsend. Best $15 I have spent.
Writing todos as checkable list items in markdown and hitting them off one by one and tracking notes on the same md under different headings works like a charm. No more JIRA / excel / context switch.
I haven’t seen on how it fares on uncensored use-cases, but from what I see Q5_K variants of Mistral 7B are not very far from Mixtral 8x7B (the latter requires 64GB of RAM which I don’t have).
Tried open-webui yesterday with Ollama for spinning up some of these. It’s pretty good.
Not sure if this helps but this is from tinkering with Mistral 7B on both my M1 Pro (10 Core, 16 GB RAM) and WSL 2 w/ CUDA (Acer Predator 17, i7-7700HK, GTX 1070 Mobile, 16GB DRAM, 8GB VRAM).
- Got 15 - 18 Tokens / sec on WSL 2 with slightly higher on M1. Can think of that to about 10 - 15 words per second. Both were using GPU. Haven’t tried CPU on M1 but on WSL 2 it was low single digits - super slow for anything productive.
- Used Mistral 7B via llamafile cross-platform APE executable.
- For local-uses I found increasing the context size increased the RAM a lot - but it’s fast enough. I am considering adding another 16x1 or 8x2.
Tinkering with building a RAG with some of my documents using the vector stores and chaining multiple calls now.
Execution traces have a goroutine profile which outputs the count of goroutines as well. That can be used for an alert as well - though it would require parsing the trace output. They recently made some changes to give a structured API over trace data - maybe use that?
There is a lot of knowledge / community support around microservices already so people find it hard to gravitate automatically towards workflows (same didn’t exist for those. Same thing as GraphQL vs REST)
(This is from when I last evaluated Cadence - which is now temporal.io. The state must have changed since then.)
Workflows are not zero-cost, they have their own tradeoffs compared to microservices. State management / bootstrapping logic becomes non-trivial, execution order though easier to visualize is also slightly not deterministic, workflows are not as well suited for request-response style replies due to the latency involved in total execution etc (but I think they are great alternatives to async / background workfllows) - and shared underlying infrastructure means increased chances of SPOFs.
The state must have improved much since then. Also, adoption of anything new to require remodelling your application into a different paradigm must be worth the value delivered. For example, modular monoliths became popular because they reduced operational complexity by reducing # of pieces involved. At the time, that value prop vs effort involved was unclear to our teams IMO
In enterprise where there is poor documentation and lots of tribal knowledge, noting down just those 2 lines for every new info is a quick way to break down knowledge gaps created by just that.
It is exactly due to such disdain for documentation that most people find it hard to navigate large codebases. Documentation is not just for noting things down pedantically but also a thinking tool and a temporary thought buffer.
And no one pushes code to production to validate assumptions. Not if you have 100 clients and you are not doing CD.
I hope they either never make it to IPO or crash violently.