James Miller

AI Solutions Architect at #TrilogySoftware

Crafting conversations between business systems with AWS & AI. #TechWhisperer #MiddlewareMaestro

[James Miller @JamesMSolutions]

Going through this now...

But also finding that in the integration process, the integration is the task, and the LLM aspect can be reduced to a very small step in the process.

Is the AI quest for automation a task of unraveling business system integrations in a costume?

[Bindu Reddy @bindureddy]

Today’s LLMs can already automate most tasks

The bottleneck is integrating these LLMs into existing processes and systems

This is a slow and painful process and will take years

[James Miller @JamesMSolutions]

While I think Machine Learning and AI can lead to improvements in safety in cars, I have no idea why I would want an "emotional connection" to a car as much as I would want one to a hammer or a mop...

[Amazon Web Services @awscloud]

Are you ready to experience mobility with an emotional connection? 🤗��� https://t.co/9XzvWEzDFG

[James Miller @JamesMSolutions]

@Replit 's new Replit Agent is the big news in my feed today.

So far, AI-built apps have been quick demos, usually all built in React, with no backend or integration. Maybe an external API connected.

Is Replit Agent any different?

Architecture is the next big leap.

[James Miller @JamesMSolutions]

Fishing for AI SaaS projects? Here is a thread.

A lot of responses are uninspired (write apps, emails, etc). We know more powerful AI is on the way, but finding novel use cases and tractable problems to solve still seems to be a creative challenge, even for humans.

[Aravind Srinivas @AravSrinivas]

What would you make you pay $100/mo and $1000/mo for an AI subscription product? What should it do for you?

[James Miller @JamesMSolutions]

@TheMonkDev @rauchg This seems like the logical answer to me

[James Miller @JamesMSolutions]

@fchollet Ha, interesting to see how physical media can still eke out a market against digital in cases like this.

[James Miller @JamesMSolutions]

At #Trilogy, we had a pretty smooth path to 90% resolution for non-tech intensive customer support.

For technical products (B2B On-Prem server software), 90% is a much more difficult target to hit.

[Bindu Reddy @bindureddy]

Today, AI can automate 90% of

- customer support
- procurement
- para legal work
- tech and IT support
- HR
- data analyst
- copywriting

We don't need super intelligence; we just need to wire everything up and that's the hard part.

The good news is that it is slowly happening

[James Miller @JamesMSolutions]

Any truth to this out there, python coders?

I am trying to pick up Python skills after a long time with npm and yarn with few complaints, and package management is still foreign to me.

[AJ Stuyvenberg @astuyve]

I've never had a python experience which didn't start with fighting the package manager and dependencies for the first few hours

[James Miller @JamesMSolutions]

Like this take on healthy skepticism about AGI from LLMs.

Advanced reasoning may not be on the menu, but that doesn't mean they're not valuable.

[Daniel Jeffries @Dan_Jeffries1]

Fantastic article (https://t.co/4WAyfGictu) from a DeepMind researcher on how he uses LLMs (very similar to how I do) and why they are incredibly valuable.

A lot of folks seem to misunderstand me when I say these models can't reason or won't get us to AGI without new architectures/algos. They take that to mean that the models are useless. That is absurd.

They are very useful and valuable and we will likely get new gains out of bigger ones with more data too.

We just won't get reasoning from first principals, fuzzy reasoning, long range planning, adaption on the fly and all the things we really want.

That said, who cares?

They're still amazing.

From the article:

"Most of the people online I find who talk about LLM utility are either wildly optimistic, and claim all jobs will be automated within three years, or wildly pessimistic, and say they have contributed nothing and never will."

So in this post, I just want to try and ground the conversation. I'm not going to make any arguments about what the future holds. I just want to provide a list of 50 conversations that I (a programmer and research scientist studying machine learning) have had with different large language models to meaningfully improve my ability to perform research and help me work on random coding side projects.

Among these:

* Building entire webapps with technology I've never used before.

* Teaching me how to use various frameworks having never previously used them.

* Converting dozens of programs to C or Rust to improve performance 10-100x.

* Trimming down large codebases to significantly simplify the project.

* Writing the initial experiment code for nearly every research paper I've written in the last year.

* Automating nearly every monotonous task or one-off script.

*Almost entirely replaced web searches for helping me set up and configure new packages or projects.

*About 50% replaced web searches for helping me debug error messages

If I were to categorize these examples into two broad categories, they would be 'helping me learn' and 'automating boring tasks'."

My number one take away from this?

Ignore the extremes.

Life is always somewhere in between.

Right now LLMs are very useful but they are not God like and magic will not spring forth from their digital weights.

The best thing you can do with AI (and life), is keep your head when all others are losing there. File the extreme views off of both sides (positive and negative) and right in that middle range is where reality lives.

[James Miller @JamesMSolutions]

Adapting whisper for improved performance. With a very clever name!

[Rohan Paul @rohanpaul_ai]

New ultra-fast ‘multi-head’ speech recognition model drop from @_aiOla, beats OpenAI Whisper.

Officially dubbed Whisper-Medusa, the model builds on Whisper but uses a novel “multi-head attention” architecture that predicts far more tokens at a time

So they seem to have added more attention heads on top of whisper. They claim the same accuracy but 50% faster. Their demo does one text in 1.9s while "baseline" whisper does it in 4s. Code and weights opensource under MIT.

they have started with a 10-head model but will soon expand to a larger 20-head version capable of predicting 20 tokens at a time, leading to faster recognition and transcription without any loss of accuracy.

[James Miller @JamesMSolutions]

Saving for later. Speech-to-text is an issue for my team that we are currently vendor-locked for, but good to keep other possibilities in mind.

[Satya Mallick @LearnOpenCV]

Enhance Speech Recognition in Air Traffic Control with Whisper. We've tailored OpenAI's powerful Whisper model to better understand the unique challenges of air traffic communications.

Discover how customizing Whisper can improve safety and accuracy, with insights into our methods and results.
https://t.co/V6Sz7Ce1SG
Dive into the specifics of our dataset, fine-tuning techniques, and the impactful outcomes on error rates and processing speeds.

#SpeechRecognition #WhisperModel #AirTrafficControl #TechInnovation #AI #asr #openai #whisper

[James Miller @JamesMSolutions]

Interesting discussion ensues here on syntactic sugar vs. type safety in TS vs. run time validation. #typescript

[David Cramer @zeeg]

why is JSON.parse<TypeStruct> not a thing in TS?

[James Miller @JamesMSolutions]

@rubenhassid Where is the result of the HTML and CSS generated? That seems like the main part of the test.

[James Miller @JamesMSolutions]

@jxnlco Worth thinking about the best forms for GenAI going forward. Not everything is a chatbot. It's usefulness is not just the content it produces, but how we consume it

#GenAI #UXDesign

[James Miller @JamesMSolutions]

Much more relaxed during a ChatGPT outage rather than an OpenAI API outage #genaideveloper

[Ray Wong @prompt_mastery]

Yes, ChatGPT is currently down.

[James Miller @JamesMSolutions]

Funny code snippet that explains the underworking of prompt optimization pipelines. The comments call it a simplified #dspy

[jason @jxnlco]

thats all

[James Miller @JamesMSolutions]

Handy tool for #GenAICoding for wide context issues and fixes

[Advait Raykar @AdvaitRaykar]

@SullyOmarr @LockerzLBR @NickADobos @vercel https://t.co/jN58ePbtYm

For others who also couldn't find it on the thread.

[James Miller @JamesMSolutions]

I have seen some complaints that #gpt4o is not as large as jump as some would like with reasoning. For half the price, how much can we complain?

[Sam Altman @sama]

in the API, GPT-4o is half the price AND twice as fast as GPT-4-turbo. and 5x rate limits.

[James Miller @JamesMSolutions]

Hah, didn't know this trick of asking for #mermaidCode for diagrams. This might help me with some cleaning up some READMEs.

[Abhishek @HeyAbhishek]

ChatGPT can now create architecture diagrams.

No need to spend endless hours making diagrams for your projects or presentations.

Here’s how to create it for free in just a few seconds:

[James Miller @JamesMSolutions]

Interesting use cases to keep in mind. The instant dashboard is a cool one.

[Bindu Reddy @bindureddy]

Here are what people are doing with AI agents today -

Email automation, doc generation, on the fly dashboards, plot generator, script and code generation!

You can do complex word docs, generate simple ppts and answer questionnaires

You can’t replace someone’s job completely but you can improve productivity by almost 50% in some cases!

If you are realistic about Gen AI, it can be instantly useful in small ways