Luke Markham

ML Engineer @ Graphcore

[Luke Markham @LukeEMarkham]

@jomatech you inspired me to create DishwasherBench!

Cheers for solving Navier Stokes along the way, really helped me ;)

https://t.co/jvf9cjiQo4

[Luke Markham @LukeEMarkham]

AI evolutionary tree 🌴🤖

🎯The emergence of LLM prompting was surely a milestone in the democratization of AI. LLMs could have easily been powerful, but hard to operate; indeed many of the early models were designed to perform a specific task that was unconfigurable at inference time. Examples are sentiment-analysis, part-of-speech tagging, sentence classifications etc.

First proposed in a 2009 paper from Palatucci et al, zero-shot inference - the idea that models could perform on tasks not seen during training - evolved and was popularized in LLMs through generative pretraining with GPT-1 and decoder-only models.

#programming #chatgpt #ai [arXiv 2304.13712 Yang et al.]

[Luke Markham @LukeEMarkham]

5 years ago I might've dismissed a new framework as simply a UX improvement and not something that'll advance SOTA

Today I'd probably describe abstractions as one of human's greatest inventions

Better abstractions are certainly needed in the prompt engineering field

#ChatGPT

[Composo @composo_ai]

🧵it feels like today's LLM applications only scratch the surface of what's possible with today's LLMs

we're starting to see empirical results that suggest the more tokens spent working through a problem the more likely the output is to be intelligent
#ChatGPT #programming

[Luke Markham @LukeEMarkham]

💾😱 using the chatgpt research preview today is like buying a server and hosting a html webpage in 1992 👇

📱apps, IM and social media only became possible because of the boom in the mid to late 90s which brought javascript, server-side programming, cookies, and more recently web frameworks, design tools and mobile networks

LLMs are the core technology, but it will take years to create the ecosystem application developers need to build the killer app

what do you think are the first fundamental technologies to emerge? model triage and tool usage perhaps? 🤔

#AI #chatgpt #llm #promptengineering [Qin 2023 2304.08354]

[Luke Markham @LukeEMarkham]

@AmplifyWithAI Good point, it does feel like it would be getting towards AGI or superintelligence - humans probably struggle with this as much as today's AIs do!

[Luke Markham @LukeEMarkham]

soooo what I’m taking from this is there’s a gap in the market for models that excel at tasks where the required knowledge is inconsistent with the real-world 😵‍💫💫

#AI #ChatGPT #LLM #promptengineering [Yang 2023 2304.13712]

[Luke Markham @LukeEMarkham]

releasing #ai applications without proper testing - especially adversarial testing - would be downright irresponsible 🤔

#ChatGPT

[Composo @composo_ai]

1/6: 🌊Calling all #AI aficionados! Who's ready for a deep-dive into language model evaluation? We're hitting a high note with 4 game-changing elements! 🚀 #PerformanceTesting

[Luke Markham @LukeEMarkham]

Pro-tip: Tired of multi-hop prompts not returning the right result? Split your multi-hop LLM prompt into multiple single-hop ones.💡

Better performance and much easier to programmatically intervene. What do you think?

#AI #ChatGPT #LLM #promptengineering [Sakarvadia 2023 arXiv:2309.05605]

[Luke Markham @LukeEMarkham]

@AmplifyWithAI Yea. There’s a good rabbit hole to go down on how decoder only models such as GPT3 produce embeddings if you’re super keen.

[Luke Markham @LukeEMarkham]

OpenAI's GPT-4 may rule the LLM league, but when it comes to embedding models, open-source alternatives like E5 and Instructor not only outperform on evaluation benchmarks 📊 https://t.co/pzZ16EwB0G but also and provide more customization.

[Luke Markham @LukeEMarkham]

It's no coincidence #LLMs are the first #AI developed that matches human level performance. Text is an efficient, information-dense format.

More data-intense formats like image and video will guarantee AI labs will be #GPU constrained for at least the next decade.

[Luke Markham @LukeEMarkham]

LLM output validation slowing your response times too much? Consider using unvalidated output and validating it post-hoc. If you control the UI you can easily recall it once validation finishes and in that time your user probably won’t have had time to screenshot and tweet it #ai

[Luke Markham @LukeEMarkham]

Read me here: https://t.co/djjgnjK9vJ

[Luke Markham @LukeEMarkham]

In the words of @hwchase17: "The biggest mistake usually is that it doesn't retrieve the right information, but retrieval is super important and if it retrieves the right information or if it retrieves partial information obviously it's gonna get it wrong"

[Luke Markham @LukeEMarkham]

@easyJet Thanks for the quick reply!
This directly contradicts EasyJet's previous guidance though: https://t.co/hviBsooDYL
You said: "If your small sports equipment is within 15 kg and fits within a hold bag of 275 cm, then there will be no issue this being a hold luggage."

[easyJet @easyJet]

@Timobru Small sports equipment is up to 20 kg, and we don't offer this as a hold luggage weight anymore. If your small sports equipment is within 15 kg and fits within a hold bag of 275 cm, then there will be no issue this being a hold luggage. Charlotte

[Luke Markham @LukeEMarkham]

@glastonbury Someone should calculate the carbon footprint of the devices and servers that go into this mess. A balloting system would be much more environmentally friendly.

[Luke Markham @LukeEMarkham]

@Iberia_en absolute scammers. Not telling passengers their flights are actually cancelled so they think they’re not entitled to a refund. Why did I have to phone you to find this out - not enough staff is not an excuse because you’ve offered me vouchers plenty of times.

[Luke Markham @LukeEMarkham]

@Gatwick_Airport I’m flying from Nantes with a systems engineer. How about you let our plane land and we’ll get your control systems working before dinner time. We got a deal?

[London Gatwick LGW @Gatwick_Airport]

UPDATE: 10 July 18.10: Due to an air traffic control systems issue in Gatwick’s control tower, flights remain suspended. We are working with ANS, our provider, to rectify this issue as quickly as possible. We apologise and advise passengers to check flight info with your airline.