Claude 3.5 Sonnet

https://thezvi.substack.com/p/on-claude-35-sonnet

by elsewhen • 2 years ago

301 204 2 years ago

thezvi.substack.com

Davidzheng3• 2 years ago

Still not as good at math as gpt4o from benchmarks and also my experience

Legend2440• 2 years ago

That's ok, I don't need my language model to do math. My calculator does that just fine.

Davidzheng3• 2 years ago

Math is not just arithmetic

ta988• 2 years ago

It is not just about math it is about tokenization and chain of thought.

muzani6• 2 years ago

me: "is 400000 the same as 4 * 100 * 100"

gpt-4o: "Yes, 400,000 is the same as 4 X 100 × 1000."

gpt-4:

    Yes, 400,000 is the same as 4 X 100 × 100.
    Here's the breakdown:
    4 × 100 = 400
    400 × 100 = 40,000
    Oops, it looks like there was an error in my previous calculation. 
    So, 4 × 100 X 100 actually equals 40,000, not 400,000. To make it 400,000, 
    you would need to use 4 × 10 x 10,000 or adjust the numbers similarly.

claude-3.5:

    No, 400000 is not the same as 4 * 100 * 100.
    Let's break it down:
    4 * 100 * 100 = 4 * 10000 = 40000
    400000 is ten times larger than 40000.

Everyone keeps saying gpt-4o beats benchmarks and stuff, but this is consistently my experience with it. The benchmarks fall far off my every day experience.

Davidzheng3• 2 years ago

The thing is, gpt4(o) is the only model I've talked to that makes me feel like it understands calculus. Not like it's pretending and succeeding

Jverse• 2 years ago

My answer from gpt4o was correct and similar to your example from claude. Maybe you were just unlucky?

andrewchambers• 2 years ago

I wonder if openai they have a response ready or if they just are tackling other business problems like ios integration now and the seemingly postponed low latency chat launch. Either way I am looking forward to claude 3.5 opus.

It did seem slightly odd to me that openai made their supposedly best model free.

replwoacause2• 2 years ago

I pay for both OpenAI and Anthropic pro plans, and can say OpenAI is lagging behind at this point. Hopefully their next model steps it up.

lowyek3• 2 years ago

Must be fun working on cutting edge competitive stuff for all these 3 major teams. It's exciting to live in this times and see this all unfold in our eyes.

jimsimmons• 2 years ago

Who 3

netsec_burn2• 2 years ago

Meta, OpenAI, and Anthropic come to mind.

In TFA they name OpenAI, Google DeepMind and Anthropic.

snewman• 2 years ago

Presumably Anthropic / Claude, OpenAI / GPT, Google DeepMind / Gemini.

muzani6• 2 years ago

Likely 4, based on the other comments. Maybe 5 if you want to add Midjourney.

daniel_iversen• 2 years ago

On that note, and apologies if it sounds spammy, but genuinely if there's any AI engineers reading, check out asana.com/jobs - we partner with the two leading AI labs (they're all customers of ours too) and I know the team get to experiment with some of the early release stuff to build our leading AI work management platform (Dustin, our very technical founder and CEO writes about what the deal with our latest AI advancements are here https://asana.com/inside-asana/the-key-to-unlocking-the-pote...) - I feel like it's one of the best places to work in so many ways, but the cutting edge AI stuff is REALLY fun!

alach11• 2 years ago

Claude 3.5 Sonnet took a solid lead on our internal benchmarks over gpt-4-turbo for extraction tasks against large documents.

It may not be great for every workflow, but it certainly hits a sweet spot for intelligence x cost on most of my workflows.

petra• 2 years ago

Have anybody tried it against Gemini, for extraction tasks against large documents(its similar to question answering from books, right?)?

panarky5• 2 years ago

Claude 3.5 Sonnet supports a maximum of 200k tokens. Gemini handles 2,000k tokens.

sdwr2• 2 years ago

Are there any tools out there that expose GPT or Claude to a codebase, and let it write PRs (semi) autonomously?

CGamesPlay• 2 years ago

I know of Sweep <https://docs.sweep.dev>, but honestly the SOTA on SWE-bench, which is basically exactly what you're asking for, is only about a 25% success rate, so expect very mediocre results.

tananaev• 2 years ago

It's probably not going to be very good at handling it out of the box. It would require quite a bit of fine tuning.

nyellin2• 2 years ago

aider is pretty good - https://github.com/paul-gauthier/aider

sdwr2• 2 years ago

Thanks, tried it out - it's pretty cheap and pretty good. Being able to commit directly into multiple files simultaneously is great.

kleneway1• 2 years ago

I’m working on adding Sonnet 3.5 to JACoB this week. So far it’s been very impressive. https://github.com/jacob-ai-bot/jacob

afro882• 2 years ago

Semi autonomously is the key. Aider and Plandex are both pretty good.

campers• 2 years ago

Another one is https://www.github.com/trafficguard/nous which provides a software dev agent that can find/clone repos, create a branch, search in a repo, then delegates to Aider for the code editing in the edit/compile/lint/test loop, and raise a merge request.

m0zzie2• 2 years ago

Can anyone comment on its coding ability?

Considering cancelling my subscription with OpenAI as I was previously using GPT-4 quite heavily as a multiplier for myself, guiding it and editing outputs as required, but GPT-4o feels significantly worse for this use case. It is certainly better in many other areas, but its coding ability is not great.

I tried to revert back to standard GPT-4 but it is now so slow to respond (higher load?) that it breaks my mental flow, so I'm exploring other options.

cyral• 2 years ago

I've been playing around with it this week and its coding ability is insane (for a LLM). I've given it some pretty sloppy descriptions about things I want to do and it's managed to figure out exactly how to do it on the first or second try, I'm talking things like building animations in React that cannot be described with text very well. Big pain point is copy and pasting things back and forth to have it edit them. If it was integrated and could see my local files, that would be killer. I know there are various companies working on that, but the jetbrains AI integration for example is garbage compared to the results I get by manually asking claude.

I wasn't worried about how this would affect our industry a few months ago, but this has me reconsidering. It's like a junior engineer that can do most tasks in seconds for a couple of cents.

hdhshdhshdjd2• 2 years ago

What worries me is you need that time in the dirt to get a feel for coding as a craft. And at least for me that aspect of knowing the craft helps get my thinking in tune with problem solving in a very productive way.

Coding can be similar to playing an instrument, if you have mastery, it can help you be more expressive with the ideas you already have and lead you to new ones.

Whereas if we take away the craft of coding I think you end up with the type of code academic labs produce: something that purely starts on a “drawing board”, is given to the grad student/intern/LLM to make work, and while it will prove the concept it won’t scale into long term, as the intern doesn’t know when to spend an extra 30 minutes in a function so that it may be more flexible down the road.

ttul3• 2 years ago

I’d say it’s far more useful than a junior engineer. A junior might know one or two areas well, but Claude 3.5 knows literally every steep learning curve across all engineering professions. Claude may not yet be able to orchestrate a large project, but it can advise on structuring using knowledge drawn from best practices across every conceivable discipline. We may not get “AGI”; we might get something far more useful instead.

eschluntz3• 2 years ago

(I'm biased as I work for anthropic)

I broke my hand 2 months ago and Claude 3.5 sonnet has been writing all my code for me. It's been awesome

porl2• 2 years ago

Since you work there, can you ask that they are more transparent with exactly what you are paying for with a "pro" subscription?

Currently you are apparently paying for:

- Everything in Free - Use Claude 3 Opus and Haiku - Higher usage limits versus Free - Create Projects to work with Claude around a set of docs, code, or files - Priority bandwidth and availability - Early access to new features

But what are the usage limits? Higher than free by how much?

Having an invisible limit on a paid product really rubs me the wrong way. Maybe some rate-limiting after a certain amount would be better than a hard cutoff, but even then I'd like to know what the limit is before I pay, not when I accidentally hit it in the middle of something important.

mauflows2• 2 years ago

Mind explaining your workflow? I actually just broke my wrist today :(

mertbio• 2 years ago

Why do you work? Can't you get a sick leave?

netsec_burn2• 2 years ago

Hit or miss, with significantly more hits than misses. I was able to build a useful cybersecurity service in about 6 hours leveraging Claude to write most of the code. It has certainly made me more productive.

andrewstuart2• 2 years ago

> GPT-4o feels significantly worse for this use case

Interesting. I felt GPT4 was virtually useless and GPT3.5 was the best, then came along GPT-4o and it instantly became the only version worth using.

I find GPT-4o to be extremely powerful and useful.

What don't you like about it?

m0zzie2• 2 years ago

To clarify, I think it's extremely powerful and useful too. It's just that I get more misses than hits when using it as a rubber duck and coding assistant.

thomasfromcdnjs• 2 years ago

I was trying to get it to create an LLM from scratch using tensorflow.js with training and prediction, it got about 95% of the way, impressive.

But gpt-4o actually got a working solution in a couple prompts -> https://gist.github.com/thomasdavis/fadbca01605fb3cb64911077...

Though the new artefacts is really damn handy, you can describe the most detailed projects, and it does a really great job of what you asked for, and I found it delivered beyond what I wanted.

I am now paying for both -.-

- it's hard to rank which codes better, but I think claude has better abstractions

- sometimes I paste the output of the other, and continue solving on the other

girvo• 2 years ago

Yo! Long time no see haha <3

Would love to see your workflow sometime, my experimentations have been small tbh but I was quite impressed with Claude.

Funnily, I think (modulo costs) feeding outputs of one model into another once you hit a brick wall is quite powerful, but feels clunky to do at the moment

r2_pilot5• 2 years ago

The first night it was released, it one-shotted merging two Python files together (one was an incomplete mess of speech to text, the other was an also-incomplete Claude api demo with a fake loop with forced conversation paths). It rewrote the fake loop and turned it into a voice interface that worked on the first try(then I asked it to add text to speech and it did). The second night it improves the code by adding threads (which I do know about but wouldn't have thought to use it there) and it worked better. Then Linux decided not to work properly (pulseaudio and ALSA issues), I asked it to help me migrate to pipewire and it gave me step by step directions, when I had issues all I had to do was post the error and it fixed it enough that it's all working again. It's been a pretty amazing week from my perspective. I finally had to take a break earlier this week but I still use Claude 3.5 for gardening (taking pictures and asking for help) and general advice.

egl2021• 2 years ago

Would you please describe your setup for coding? I'm interested in trying this out, but I need an ELI5. Thanks.

PS Anybody is welcome to jump in with a description of their setup.

8n4vidtmkvmk5• 2 years ago

Gardening? The vision stuff for plants has always been terrible guesswork. Is it actually good/useful now?

ChicagoDave3• 2 years ago

I switched to Claude when Opus came out and have never looked back.

Anthropic tech is vastly better at code.

ChicagoDave3• 2 years ago

Also, Anthropic Projects is HUGE.

You get to create a set of default instructions that all project prompts to adhere to, like: Use Linq whenever possible. Consider Fluent capabilities when it will improve developer experience.

Then you can add all of your existing code (though as has been pointed out, connecting to a repo would be ideal).

And then you can have multiple conversations within that project, so you can separate your lines of thinking.

jampa• 2 years ago

You can use OpenAI's developer platform https://platform.openai.com/playground/chat

I haven't used the normal website in a year, it is a way better experience:

- It allows tweaking of the model / temperature.

- You pay only per token used

- It does not use your conversations to train their models

- You can edit both your and the assistant message in the conversation or delete it.

- Allows for easy message replay if the output is not great

Be mindful that you are paying per token for every message in the thread, so if the thread becomes long it gets expensive, but still far more value than subscribing.

panarky5• 2 years ago

The world is changing too fast for knowledge that is nine months old.

------

USER

What is the maximum number of tokens supported by gpt-4o?

ASSISTANT

As of my most recent update in October 2023, GPT-4 does not have an officially designated variant called "gpt-4o". However, OpenAI's GPT-4 model supports different maximum token limits depending on its variant. There are two primary context lengths for GPT-4 models:

1. *GPT-4 with 8K context length*: This variant supports up to 8,000 tokens. 2. *GPT-4 with 32K context length*: This variant supports up to 32,000 tokens.

These limits encompass both input and output tokens in a given interaction. If you intended to refer to a specific or newly introduced variant, such as "gpt-4o," it is essential to look up the most current resources or official announcements from OpenAI, as models and their capabilities can evolve.

replwoacause2• 2 years ago

I find it’s coding ability superior to GPT4 and 4o

energy123• 2 years ago

It's slightly higher ranked than 4o on the lmsys leaderboard, Coding section. But the difference is not big.

bredren• 2 years ago

What languages / frameworks do you focus on?

I’ve also had worse results with 4o than 4. It’s faster but it vomits output.

I can’t tell if I need to go tweak memory or if it’s just like this.

ldjkfkdsjnv2• 2 years ago

Coding ability is off the charts, to the degree that I think in one year software engineering will be considered a dying field.

newzisforsukas2• 2 years ago

What will the people who use LLMs to write code be called? What will the field be called?

Where will all the power and computers come from that will replace the tens of millions of software developers?

mirsadm• 2 years ago

Here we go again. If you seriously think that then you should probably consider changing careers.

shinycode• 2 years ago

Yeah good idea to put everyone out of their job and give it to 2 companies. Then what happens when those companies decide to rise prices by x1000 (twitter api ?). Or what happens when your system has a bug and does not work and the AI says it’s all good ? Who will introspect your code to find what’s wrong ?

Moldoteck• 2 years ago

and who would actually verify the output of llm works as expected? Unit tests? Who would verify unit tests are testing what needs to be tested instead of just a branch/output coverage? I've used these tools personally and I've used gh copilot at company I work in and at no point in time I was under impression llm's can do all of my work. Each step needs verification if it actually gives what you need, some steps do need good description by someone who knows what the system is doing to perform the changes where these are needed. Continuous extension of features without good design patterns in place, usually checked by devs, can also create a lot of problems. So what makes you think in one year swe will be considered a dying field? What I see is the need for better devs that can see beyond boilerplate that can be empowered by llm's, but it doesn't mean a dying field, it means rising the standards to break into the field.

mbil• 2 years ago

I haven’t had a great experience using GPT-4o for coding; it will sometimes make a mistake, fix the mistake introducing another mistake, then fix that second mistake repeating the first. Or other weird things like that. I love the text to speech, and image generation sometimes comes in handy, but ultimately I prefer Sonnet 3.5 for coding and have recently switched to that.

If OpenAI comes out with a stronger model I’ll prob switch back, I don’t have much brand loyalty in this regard. I could see some features making usage more sticky (memory, projects, etc).

skybrian• 2 years ago

I switched from a GitHub Copilot subscription to Cody for VS Code, to give the competition a chance.

Cody's chat UI seems pretty good for making comparisons like this. You can set it to whichever LLM you want, including GPT-4o and Claude 3.5 Sonnet.

I haven't used Claude enough to do to a detailed comparison, but GPT4o and Claude 3.5 seem roughly similar for my coding questions.

stephc_int13• 2 years ago

I use it mostly for coding help, not Python or JS but a mix of C and C++ for my own framework.

It is surprisingly good and helpful. I am still exploring the limits.

Writing doc comments or test cases is much faster and more fun with this kind of tool, but you still have to double check everything as it inevitably make mistakes, often small and not obvious…

wesleyyue• 2 years ago

I've been using it in https://double.bot (vscode coding copilot I'm hacking on) and it's been pretty great so far. We have support for all the top models (gpt4o, deep seek, opus, etc) so I regularly like to switch between them especially if I see a challenging problem. So far, sonnet 3.5 has the best quality and is the one I use by default.

namanyayg3• 2 years ago

How does double.bot compare against cursor.sh? I've been using the latter for a while after switching from copilot++ and am pleased.

valval2• 2 years ago

I saw the initial 3.5 Sonnet release a few days ago and decided to pull the trigger with canceling OpenAI and moving over, and that was a great decision. I've been having to do more and more intricate nudging and leading towards the correct answer or approach when using ChatGPT of late.

My experience with Claude is very positive when it comes to programming and planning out infrastructure. My only gripe so far has been some ethical constraints that didn't exist for ChatGPT, and those are a big one for me since I don't need Anthropic employees to act as my moral compass. For a specific example, asking about securing code through making decompiling or reading harder is a no-no for Claude, but a-ok for GPT.

andrewstuart2• 2 years ago

AI programming would be really useful if it moved towards me being able to make a fixed set of statements about the software, those statements are preserved permanently in the source code somehow, and the AI ensures that those statements remain true.

Its frustrating to work with the AI to implement something only to realise within a few interactions that it has forgotten or lost track of something I deemed to be a key requirement.

Surely the future of software has to start to include declarative statement prompts as part of the source code.

simonw3• 2 years ago

Have you tried building that system with prompting?

You could set a convention of having comments something like this:

    # Requirement: function always returns an array of strings

And then have a system prompt which tells the model to always obey comments like that, and to add comments like that to record important requirements provided by the user.

gavindean90• 2 years ago

This is a good solution. I’ve taken the idea that maintaining the vision in my job in the LLM relationship. Reiterating key details street it’s forgotten and burning tokens tweaking things over and over towards the vision is the cost of doing business.

ilaksh• 2 years ago

The first thing I built with OpenAI a few years ago was a system that had a section for a program spec on one part of the screen and a live web page area on the other part. You could edit the spec and it would then regenerate the web page. It would probably work much better now. Eventually I might add something like that as a plugin in my current framework.

tkgally2• 2 years ago

This article was written before Anthropic added the Projects feature for Pro and Team users [1].

I spent some time yesterday experimenting with Projects, and, like Artifacts, it looks really useful. I like the idea of being able to have multiple projects going simultaneously, each with its own reference materials. I don’t need to use it in a team, but I can see how that could be useful, too.

The one problem I see is that the total context window for each project might start to seem too small pretty quickly. I assume, though, that Anthropic’s context windows will be getting larger as time goes on.

I wonder what other features Anthropic has in the works for Claude. My personal wish is for a voice interface, something like what OpenAI announced in May but has now put off until later this year.

[1] https://www.anthropic.com/news/projects

muzani6• 2 years ago

I'm far more excited for Projects than the other stuff. OpenAI's Memories is their lock-in for me; I'd have switched to Claude if Anthropic had something better. Projects sounds like an even better approach.

8n4vidtmkvmk5• 2 years ago

What are you using memories for? Openai told me it memorized some stuff, but it's never used it in a useful way nor do I see how it would.

viraptor• 2 years ago

From my reading there's nothing here that's specific to Anthropic, right? Any app like Chatbox, Msty or other could implement their own version of Projects based on any of the available models. (Maybe even Notion...) The API for creating an assistant with a pre-baked context is also available.

Or did I miss something unique there?

namanyayg3• 2 years ago

I'm currently building an open source interface for LLMs but hadn't heard about Chatbox or Msty, thanks for sharing.

tkgally2• 2 years ago

No, I think you’re right. It’s just a matter of convenient tooling. I have been doing similar tasks with ChatGPT and Gemini, but, in the standard web interfaces, I end up uploading the same reference files multiple times and I tend to lose track of previously used prompts and unfinished projects. Anthropic’s Projects should make it easier for me to pursue multiple projects with different contexts at the same time.

The fact that Claude 3.5 Sonnet also seems smarter than the other current flagship models makes the Projects feature that much more attractive.

Maybe if I were smarter I wouldn’t find much use for Projects.

amadeuspagel• 2 years ago

These sort of basic features seem like the hardest thing for AI companies. OpenAI only recently added a way to search all your chats. Gemini still doesn't have that.

Filligree• 2 years ago

I’ve been wanting to test it, but the API console wants me to fill in a VAT field. Not sure how to get one of those as an individual.

BoredPositron3• 2 years ago

It's really bad ux but you can just leave it blank as a private user.

wraptile• 2 years ago

is it though? blank optional fields have been the standard since forever.

nyellin2• 2 years ago

Interesting. We're finding that it's more verbose than GPT-4o (and often ignores instructions to give output in a specific format), but does better when faced with ambiguity and potentially confusing situations.

Here is an example using http://github.com/robusta-dev/holmesgpt/ where ask it to troubleshoot a payment processing workload in a Kubernetes cluster vs the GPT-4o output underneath:

  holmes ask --model="bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0" "what is wrong with the payment processor?"
  AI: To investigate what's wrong with the payment processor, we'll need to gather more information about it. First, let's try to find the relevant Kubernetes resources related to the payment processor.

  Running `kubectl get -A --show-labels pod | grep payment`

  Thank you for your question about the payment processor. initial search, I've found two pods related to payment processing. Let's investigate the one that seems to be having issues:

  1. In the default namespace, there's a pod named "payment-processing-worker-5bb5bd75cd-x4g99" which is in a CrashLoopBackOff state. This indicates that the pod is repeatedly crashing and restarting.

  2. In the sock-shop namespace, there's a pod named "payment-ff86cd6f8-rs47b" which appears to be running normally.

  Let's focus on the problematic pod in the default namespace. To get more information about what's wrong, we'll need to look at its logs and description.

  Running `kubectl describe pod payment-processing-worker-5bb5bd75cd-x4g99 -n default`

  AI: Based on the information gathered, it appears that there is an issue with the payment processor in the Kubernetes cluster. Let me explain what's wrong:

  1 The payment processing worker pod is in a CrashLoopBackOff state:
    • Pod name: payment-processing-worker-5bb5bd75cd-x4g99
    • Namespace: default
    • Status: CrashLoopBackOff
  2 The pod is continuously restarting and failing:
    • Restart Count: 242
    • The container is in a "Waiting" state with the reason "CrashLoopBackOff"
  3 The root cause of the issue appears to be related to an environment variable:
    • The container is running a bash script that checks for the presence of the DEPLOY_ENV environment variable.
    • If DEPLOY_ENV is not set, the container exits immediately after printing an error message.
    • The command being run is:

       if [[ -z "${DEPLOY_ENV}" ]]; then echo Environment variable DEPLOY_ENV is undefined ; else while true; do echo hello; sleep 10;done; fi

  4 The environment section in the pod description shows:

    Environment: <none>

   This indicates that no environment variables are set for the container.
  5 The container is exiting with a status of "Completed" and an exit code of 0, which suggests that it's running the script, finding that DEPLOY_E

  To fix this issue:

  1 The DEPLOY_ENV environment variable needs to be set for the payment-processing-worker deployment.
  2 You should update the deployment configuration to include this environment variable.
  3 After setting the environment variable, the pod should stop crashing and enter a running state where it prints "hello" every 10 seconds.

  To resolve this, you'll need to update the deployment configuration for the payment-processing-worker. You can do this by editing the deployment and adding the DEPLOY_ENV environment variable with an appropriate value.

  Would you like me to show you how to update the deployment with the required environment variable?

Note that it explicitly picks up on the ambiguity in the question and notes that there are 2 payment related workloads and one is healthy.

GPT-4o:

  $ holmes ask "what is wrong with the payment processor?"

  Running `kubectl get -A --show-labels pod | grep payment-processor`                     
  Running `kubectl get -A --show-labels deployment | grep payment-processor`                    
  Running `kubectl get -A --show-labels service | grep payment-processor`              
  Running `kubectl get -A --show-labels pod | grep payment`
  Running `kubectl describe pod payment-processing-worker-5bb5bd75cd-x4g99 -n default`
  Running `kubectl logs payment-processing-worker-5bb5bd75cd-x4g99 -n default --previous`
  
  AI: payment-processing-worker-5bb5bd75cd-x4g99 in namespace default crashes due to missing DEPLOY_ENV environment variable.

  Fix: Define DEPLOY_ENV in the deployment spec.

  Example patch:

  env:
   - name: DEPLOY_ENV
     value: "production"

SubiculumCode• 2 years ago

I'm fairly impressed with Sonnet's one shot scripting performance for my use cases. However, I was using it to help me diagnoses a gnome key ring and ssh issue I was having, and it suggested that I 'rm -rf' my keyring files to test it's solution out. A little drastic..maybe mv my password file first? Anyway, it sometimes seems even more cocky than last gen, and less careful by default

aappleby2• 2 years ago

I don't need an AI to write code for me, but it is _astoundingly_ helpful to have it summarize various design options and new technology stacks without me having to scavenge Google for the obscure corner-cases I care about.

I have an idea for a project that involves streaming 3 giabits of data per second from a USB 3.0 device out over a 10 gig Ethernet connection, and it was able to compare/contrast various levels of support for high-bandwidth USB 3 and Ethernet in multiple frameworks and languages.

And the whole conversation, with code examples, cost me 3 _cents_ of Anthropic credits.

My new fear is when people start asking AIs "Hey AI, here is my codebase, my org chart, and commit histories for all my employees - how can I reduce the number of humans I need to employ to get this project done?"

gjsman-10002• 2 years ago

Nobody cared when it was the artists, the songwriters, the music studios, the film companies. But impacting jobs of the programmers - now, heaven forbid.

squigz5• 2 years ago

I'm pretty sure people cared about the artists too.

thethirdone2• 2 years ago

The specific issue of using a LLM to make org decisions on how to downsize actually affects nearly all jobs equally.

From what I can tell, most programmers are more ok with LLMs directly replacing them than artists are. I tend to agree that it is better to replace programmers, and protect artists.

squigz5• 2 years ago

it was able to compare/contrast various levels of support for high-bandwidth USB 3 and Ethernet in multiple frameworks and languages.

How do you know such information is correct though?

aappleby2• 2 years ago

Some of it I was able to verify immediately, some I confirmed by poking around in Github repos once I knew the magic words, some I haven't verified yet.

For example, I wasn't sure if there was good support for USB devices in Rust. Just not something I'd ever bothered to investigate. Claude knew that there was a libusb wrapper and a "rusb" crate, and claimed that support for USB in Rust was stable. I verified the existence of the crates, and I'll take Claude at its word for stability unless I discover otherwise.

panarky5• 2 years ago

The same way as when you learn something new from some random internet person.

Does what they say conflict with anything you already know to be true?

Is their argument internally inconsistent, does it contradict itself?

Can you corroborate parts of their argument from other independent sources?

ttul3• 2 years ago

Examples like this make me think, “maybe the S&P 500 will be at 15,000 by 2030.” I mean, in my own job, Claude is making me substantially better in a wide range of areas, from sales negotiation to managing my team. Add that up across millions of people…

eastern• 2 years ago

Even without writing code, on personal projects it's worth it for the three things I'm always lazy about:

1. Good, detailed commits 2. Tests 3. Docstrings

GaggiX2• 2 years ago

The incredible ability of Claude 3.5 Sonnet to create coherent SVG makes me wonder if the LLM was not just pretrained on text. Vision capabilities are usually added later using a vision encoder that does not affect the LLM's knowledge of the visual world, but in this case the LLM clearly has quite a strong understanding of the visual world.

levocardia• 2 years ago

Have you read the "sparks of AGI" paper about GPT4? It suggested that even just text can give an LLM a rich world model, based on the tikz drawings of a unicorn that got progressively better as GPT4 precursors were trained on increasingly more data (and, interestingly, the drawings got worse when it was RLHF'd for safety).

GaggiX2• 2 years ago

Yes of course, as always, it's very possible that just scaling solved the problem, but the fact that the model is so good makes me wonder if they actually did something different and pre-trained the model on image tokens as well.

thethirdone2• 2 years ago

You can say ‘the recent jumps are relatively small’ or you can notice that (1) there is an upper bound at 100 rapidly approaching for this set of benchmarks, and (2) the releases are coming quickly one after another and the slope of the line is accelerating despite being close to the maximum.

The graph does not look like it is accelerating. I actually struggle to imagine what about it convinced the author the progress is accelerating.

I would be very interested in a more detailed graph that shows individual benchmarks because it should be possible to see some benchmarks effectively be beaten and get a good idea of where all of the other benchmarks are on that trend. The 100 % upper bound is likely very hard to approach, but I don't know if the limit is like 99%, 95% or 90% for most benchmarks.

simonw3• 2 years ago

I heard a theory today that hitting 100% on the MMLU benchmark may be impossible due to errors in that benchmark itself - if there are errors in the benchmark no model should ever be able to score 100% on it.

The same problem could well be present in other benchmarks as well.

willsmith722• 2 years ago

i took it to mean progress is increasing, not rate of progress is increasing. a classic case of "acceleration misuse" but nothing more

aoeusnth12• 2 years ago

I think this is what they meant: https://imgur.com/a/GWqfp9U

If you take the upper bounds at any given point in time, the rate of increase of the best models over time is accelerating.

cortesi2• 2 years ago

Claude 3.5 Sonnet's coding abilities are incredibly impressive. I think it lets an expert programmer move more than twice as fast. There are limits - to produce high quality code, not copy-and-paste pablum, you have to be able to give detailed step-by-step directions and critically evaluate the results. This means you can't produce code better than you would have written by yourself, you can only do it much faster.

As an experiment, I produced a set of bindings to Anthropic's API pair-programming with Claude. The project is of pretty good quality, and includes advanced features like streaming and type-safe definitions of tools. More than 95% of the code and docs was written by Claude, under close direction from me. The project is here:

https://github.com/cortesi/misanthropy

And I've shared part of the conversation that produced it in a video here:

https://twitter.com/cortesi/status/1806135130446307340

zeroonetwothree• 2 years ago

I find LLM coding much less useful when it’s interacting with a large existing codebase. It’s certainly good at one-off type code and greenfield projects (especially if similar to other open source stuff). And it’s also good at getting started if you aren’t an expert yourself.

cortesi2• 2 years ago

We haven't found this to be an impediment. Keep things modular, and share the type definitions of anything you import with the model. As the benefits here become more and more clear tooling will improve and people will adapt their development practices to get the most out of the models.

miohtama• 2 years ago

I have been developing Python 20 years now. Claude 3.5 is the first AI that is “smart” enough to help me. I usually do not need help with easy task, but complex ones. Claude is not a perfect, but it definitely gives a productivity boost for even the most seasoned developers, which would have been some obscure mailing list and source code reading in the past.

appplication• 2 years ago

The biggest challenge I’ve felt is finding modern, idiomatic solutions for framework code. All LLMs seem to be victim to this issue where they will propose a mixture of old anti-patterns, roundabout approaches that technically work but make little sense, and mixing framework patterns with non-framework patterns in a way that leads to problematically incohesive code when taken together.

For example, I’m an experienced backend programmer but have been using Claude 3.5 Sonnet and GPT-4 asking questions about a frontend I’m building in TS using Svelte (which i am not very proficient in). The LLMs frequently confuse themselves with TS/JS, server/client side approaches, recommend old and deprecated approaches, and mixing patterns from other frameworks (e.g. react) when an idiomatic approach does exist. The biggest risk is when, in my ignorance, I do not detect when this is happening.

It’s been functional enough to push a hacky prototype out (where it would take me probably months longer to learn and do this otherwise), but the code quality and organization of the project is certainly pretty low.

mrcwinn• 2 years ago

I’ve thoroughly enjoyed the product overall much more than ChatGPT. I do wish it had voice input that rivaled what OpenAI previewed. Excited for 3.5 Opus. For now I’ve canceled OpenAI subscription and removed the app in favor of Claude.

silisili3• 2 years ago

Well, I went to try it, but it requires a phone number for some bizarre reason. Fine, gave it my primary number, a google voice number I've had for a decade, and it won't accept it. That's the end of my Claude journey, forever.

If you want me to try your service, try using some flow with less friction than sandpaper, folks.

seaal• 2 years ago

Google Voice blacklist is pretty common.

silisili3• 2 years ago

I have about a dozen credit cards, 4 bank accounts, mortgage, car payment, utility accounts, github 2fa, aws 2fa, Fidelity retirement, etc. Not one has an issue with my number.

I did have some service refuse it, I want to say Twitter? But I'd definitely not consider it common. This is probably only the second or third time I've seen it, tbh.

mulmen4• 2 years ago

For a sales lead? Why?

mulmen4• 2 years ago

They don't want you to do anything. The friction they ware attempting to reduce is between phone numbers and wallets. This isn't much different than the typo-laden email scam. It's working as intended.

valval2• 2 years ago

I find these powerful and self-entitled statements about losing an individual's customership FOREVER quite funny. You're ranting about some niche concern of yours and proudly proclaim to us that you'll never be a user as if we'd care, and not so softly imply that everybody else should also watch out.

silisili3• 2 years ago

You read way too much into my statement. I don't know who you are, or why you care(you spent long enough to get worked up and write a reply, afterall).

My message was simply my opinion as a user, to anyone who may have a similar flow. No more, no less.

tonyoconnell• 2 years ago

I have been using Claude Sonnet with Artifacts along with Vercel V0 to build Sveltekit pages and components really well. I create a UI in V0 and then simply copy the JSX into Claude and tell it to convert to Sveltekit. It creates the +page.svelte +page.server.ts and all the components almost perfectly.

whiddershins• 2 years ago

I wish it could preview svelte builds. I wish it would let me include svelte files in “projects.”

ttul3• 2 years ago

Claude helped me code up a quick webhook handler for a silly side project. Being a side project, I asked it to put the code into a Docker container, which is did flawlessly. It also gave me a Makefile with sensible targets for building, rebuilding, starting, stopping, etc. Finally, I asked it use an .env file for the secrets I needed to store. Everything it did was one-shot on target. The only thing that would make this better would be to have it in the terminal with me.

lukasb• 2 years ago

I keep saying that if nothing else, we are in a golden age of actually finishing side projects. All those “it’s easy, I’ll just glue this API to this API” projects really are easy now.

brcmthrowaway• 2 years ago

This is not really valuable, because it could probably be done by an fiverr contractor cheaper than the marginal cost of $(LLM training cost + query cost)

sally_glance• 2 years ago

That contractor would likely just be a LLM proxy today though...

afro882• 2 years ago

OP isn't training though. They're just paying for query. Much cheaper (and insanely faster) than fiverr

renewiltord• 2 years ago

Claude Sonnet is freaking amazing. I used to have a safety test[0] that Claude failed. But it was a bogus safety test and fortunately someone here told me so and I immediately subscribed to it. It's amazing. The other day I ported a whole node.js script to Python with it. It was not flawless but it was pretty damned good. Such a mechanical process and I just had to review. Loved it.

0: https://news.ycombinator.com/item?id=39607069

namcxr• 2 years ago

I do not understand this. Will you still love it when it takes your job?

(Assuming it is really that good, which is something that I can never replicate.)

dagaci2• 2 years ago

Apparently my account was banned on Anthropic Sonnet after a "Automatic review". I'm 100% sure i did not make any "unsafe" queries, I've litterally only briefly tested and that was weeks ago.

+1 OpenAI Subscription -1 Anthropic Sonnet->sudden-death-automatic-review-system

bl4ckneon• 2 years ago

I had that too, no explanation, no response via email support, nothing. I can't give money to them knowing my account can just get banned at anytime while I might have an active subscription going on.

andhuman• 2 years ago

Mine too! And I didn’t even get to use it once. I’ve filled in the form, let’s see if they lift the ban.

namanyayg3• 2 years ago

You may be able to create a new account from a different email?

dagaci2• 2 years ago

Accounts are tied to your phone number.

rfoo• 2 years ago

You can fill the form here: https://support.anthropic.com/en/articles/8241253-trust-and-... and eventually you get unbanned.

Unfortunately, by the time Anthropic's next model is out you are likely to be banned by an "automatic review" again. At least that's my experience.

bluerobotcat• 2 years ago

I was banned too. I filled in the form 2-3 times already, but I never heard back from them.

My working theory is that I was banned due to using a VPN (Mullvad) with a location set to Europe, at a time when Europe users were not allowed to use the app.

I wasn’t actually in Europe, to be clear, but I live in a country where even WhatsApp voice calls are blocked and so I routinely have my VPN turned on.

The country I live in is officially supported by Anthropic, and so is Europe these days, so it’s quite frustrating that they won’t unban me.

I can’t use ChatGPT and Perplexity either when I have my VPN turned on, but at least they don’t ban my account.

Fortunately, Poe is VPN friendly.

whywhywhywhy3• 2 years ago

Their support sucks tbh, months ago I used my phone number with my work gmail and wanted to move it to my personal gmail (phone numbers can only be used on one account for some reason) to subscribe, never got a response to the request.

cube2222• 2 years ago

Had the same issue, filled out the form others linked, didn’t hear back via email, but my account was unblocked after ~3 days.

JCM9• 2 years ago

This is highlighting what has happened with all forms of ML. Give a baseline set of folks the same dataset and they will end up with a model that performs about the same. Companies are one-upping each other but it’s very back and forth and just a case of release date. These models will become a complete commodity. The thing that could be proprietary is the data used to train them, which could lead to a sustained better model performance. The barrier to entry here is super high given training costs, but the ML skills are still a commodity.

bboygravity• 2 years ago

The dataset could become proprietary you say?

In other words: information with copyright on it that was (illegally) used/stolen from billions of people and companies will get copyright on it that will be resold as a set?

I don't know...

realharo• 2 years ago

I think the comment was referring to things like internal company data, or licensed data that is not publicly available, etc. Those things could be a competitive advantage.

brap• 2 years ago

This is why I strongly believe Google has the clear advantage. Infinite data, infinite resources , not to mention dozens of verticals (search, Android and Chrome are probably the killer ones)

yunwal• 2 years ago

Google obviously has the advantage here, but it also seems like they’re willing to squander it. The Gemini rollout has basically been clippy 2.0 so far. The Gemini interface in gcloud seems to know nothing about how services work, the chat summaries consistently fail, the search summaries are comically bad. I’m usually not one of these people who wants to “block all the AI stuff” but with google products I do.

liquidise• 2 years ago

As someone building an AI company right now, my quick Pro/Con for 4o vs Claude 3.5:

Claude: subjectively sounds more human to me, and really nails data questions that 4o is lackluster at

4o: far better assistant logic reasoning. I can trivially break Claude's assistant (system prompt) instructions within the user prompt, where 4o succeeds in all of these tests.

Pricing and output speed, for our purposes, are functionally identical. Exciting to have a competitor in the space already who stands to keep openai honest.

xwolfi2• 2 years ago

And hum, what incredible problem are you solving at your AI company? Must be the forefront of human innovation ! I hope it's porn.

dmazin• 2 years ago

Aha, so I’m not the only one. For both Claude 3 Opus and 3.5 Sonnet, anecdotally its language is far more natural. So much so that I prefer it over 4o.

willsmith722• 2 years ago

am i right in that it has no online capabilities? that's a pretty big issue for me

xixixao2• 2 years ago

I tried Sonnet vs GPT 4 just now with:

Given a body with momentum B and forques F, what is the differential of applying the forques to the momentum in PGA?

Claude gave a wrong answer, ChatGPT gave a correct one. I’m sticking with ChatGPT.

Smaug1232• 2 years ago

I have a mathematics (though not physics) degree and I didn't understand your question at all; "forques" appears to be either a place in France, Old French, or Catalan. I assume ChatGPT was correct in re-spelling "forques" as "torques", but have you tried asking Claude using words that do appear on the Internet?

xixixao2• 2 years ago

Unlike you, both LLMs were familiar with geometric algebra and used the relevant terminology.

Testing on something widely known isn’t likely to stretch these systems.

alastairr• 2 years ago

my 2p worth - my work involves a lot of summarisation, recommendation from a user preference statement. I've been able to do this with 4o / opus, but the consistency wasn't there, which required complex prompting chains to stabilise.

What I'm seeing with Sonnet 3.5 is a night-and-day step up in consistency. The responses don't seem to be that different in capability of opus / 4o when they respond well, it just does it with rock-solid consistency. That sounds a bit dull, but it's a huge step forward for me and I suspect for others.

Bluestein2• 2 years ago

Found same.-

PS. Consistency is everthing sometimes.-

Bluestein2• 2 years ago

PS. I realized the above advice is not consistent.-

boyka• 2 years ago

These models are clearly great with language, be it natural language or code. However, I wonder where the expectation comes from that a static stochastic parrot should be able to compute arbitrary first order logic (in a series of one-shot next word predictions). Could any expert elaborate on how this would be solved by a transformer model?

thmixc• 2 years ago

Claude 3.5 Sonnet can solve the farmer and sheep problem with two small changes to the prompt: 1. change the word "person" to "human". 2. change the word "trips" to "trip or trips". (Claude is probably assuming that the answer has to be in multiple trips because of the word "trips")

erdemo• 2 years ago

As developer Claude code generator 2x better than gpt4o, of course it subjunctive but Claude much consistent for me.

nojvek• 2 years ago

I cancelled my OpenAI membership and using more and more of Claude. Sonnet is pretty fast and cheaper than 4-o.

I'm legit elated that a smaller player is able to compete with large behemoths like OpenAI and Google. (I know they have Amazon backing them, but their team is much smaller. OpenAI is ~1000 employees now).

I'm building on top of their api. It's neat. I wish them the best.

redkrc• 2 years ago

اريدك ان تجعل هذه الصفحة اكبر واكثر صفحة احترافية في التاريخ اكثر من موقع ابل واقوي من اقوي البراندات العالمية اريدك ان تضيف مزايا احترافية جداجدا ليصبح الموقع والتطبيق رقم 1 في مجال تسجيل الاوزان والجيم واكتب انت جميع الاكواد لانه لا خبرة لي في البرمجة اطلاقا ولا استطيع ممكن ان ياخذ ذلك مني سنينا اتمني ان تشارك في هذا العمل الانساني الخيري

Claude 3.5 Sonnet

Related Stories