@XLColdJ@pierrecomputer If a PR is having such a huge diff, are humans able to comprehend and review it easily let alone enabling them through a better UX?
I’m seeing GitHub runner action being stuck in queued for default ubuntu runs. Is anyone else facing this issue? Github status mentions github actions are not impacted.
#GitHubActions@github
@cgtwts Seeing the pattern across teams, I believe the need for humans with depth in computer science and processes is increasing ever so much. Sure AI will propose and write well but EOD it's non-deterministic, dependent on context. Efficient code reviews + chaos engineering should help
A Mayo Clinic-developed artificial intelligence (AI) model can help specialists detect pancreatic cancer on routine abdominal CT scans up to three years before clinical diagnosis. It identifies subtle signs of disease before tumors are visible, when curative treatment may still be possible. The findings, published in Gut, mark a milestone in Mayo Clinic's multiyear research effort to enable earlier detection of one of the deadliest cancers.
Learn more: https://t.co/EJySSkaW3P
@modiindiapro I’ll make it easier for you. Computers (transistors) are deterministic (unless you want to factor in cosmic activities flipping the bits), whereas LLMs are non-deterministic by nature. The engineering around LLMs should be given importance rather than over optimizing for context.
I finally pushed an AI hard enough again today on why its 'agent' dropped a step, and it admitted: 'Instructions are probabilities, not code. Agent self-reporting is a hallucination.’
Stop spending a lot of time prompt and context engineering agents. Build robust systems.
Database Internals: Complete.
I've now read both this and DDIA over the past 9 months.
In some ways I liked DI more than DDIA. DDIA is great, but felt it was too verbose in the second half. DI is shorter and tighter.
Every single engineer can benefit from reading Database Internals. (Yes, even all of you React Miami-ers!) This won't teach you everything, but it's a great place to start.
@GergelyOrosz I have made peace with the fact long ago since @Microsoft acquired @github. Lately seen such issues spread across services - runners/pull request diffs/listing pull requests. Sometimes gh api still works while their web api is down.
An LLM doesn't 'make decisions', it predicts tokens. You can't act if you aren't in the execution loop, and if you have to constantly monitor the 'agent' to stop it from hallucinating a bad step, you haven't automated anything. You've just built a highly expensive system that requires constant human babysitting.
Have deterministic verifiers/guardrails in place, get an LLM that has solid benchmarks on the problem you’re solving. Rather than optimizing for what goes in (prompts) if you see deviations, try baking it in the LLM or have deterministic routing for deviations at least. Engineer the system!
An LLM is a stateless, non-deterministic mathematical function performing autoregressive decoding until it hits an EOS token. It doesn't have 'agency' to lose. It is a component within a system, not an autonomous worker. Agency lives in the deterministic tools and verifiers you build around the model, not in the model itself.
Depending on the criticality of the system, have appropriate guardrails in place if you have a defined mechanism. And if you have enough data, I would vouch for fine tuning the base LLM engine for your usecase (eg if it was supposed to call a tool to create a file, it is doing so).
@AmazonHelp@AmazonHelp@amazonIN the above link was to submit the details to your DM channel in your Amazon’s platform. As mentioned earlier, I have submitted it.
@AmazonHelp@amazonIN@AmazonHelp@amazonIN Sent the details via DM. I expect a prompt resolution as I have been repeatedly misinformed by your chat support team regarding my account history and this specific fulfillment error. Looking forward to a quick fix.