Posts by StuartGray@mastodonapp.uk
(DIR) Post #ATNCNnYXd5VjC3zr6m by StuartGray@mastodonapp.uk
2023-03-07T13:10:47Z
0 likes, 0 repeats
@TedUnderwood I'm not sure if this role has a widely recognised professional job title, but how could they miss "Machine Learning Dataset Builders"? 😂
(DIR) Post #ATaIxM4OQR2n9RiCgq by StuartGray@mastodonapp.uk
2023-03-13T20:55:41Z
0 likes, 0 repeats
@TedUnderwood Interesting observation.I haven't used AI art extensively, and never for brainstorming because so far I've not had a pressing need.However, I have used ChatGPT to brainstorm some ideas for a plot line I've been tinkering although it's a bit hit & miss, I found it very useful overall.Likely not as good as an engaged & knowledgable person, but *much* better than a well-meaning but ultimately clueless friend, and gave me some new ideas I hadn't considered.
(DIR) Post #ATcH5LrBI3llDO8a2K by StuartGray@mastodonapp.uk
2023-03-14T19:44:09Z
0 likes, 0 repeats
@TedUnderwood Nice comparison.I had similar results with ChatGPT. I found that getting it to either extrapolate, create additional content, or suggest new alternatives was a mixed bag. Sometimes it worked, sometimes it didn't.I had the distinct impression that I was running up against deliberate restrictions each time it failed, and usually required creative prompt editing to get the desired result.However, your example suggests that may just be "cover" for it's lack of context.
(DIR) Post #AUN5rcFYNI3ayhjkbA by StuartGray@mastodonapp.uk
2023-04-06T09:47:17Z
0 likes, 0 repeats
@simon LLMs can’t lie, they can only ever output tokens according to statistical probability derived from their training. It responds to its input exactly as it was trained to do with zero understanding or agency. Please don’t fall into the anthropomorphism trap like so many others.This is a great, clear read on the differences between the ways in which humans think and LLMs predict, a short paper by Murray Shanahan https://arxiv.org/pdf/2212.03551.pdf
(DIR) Post #AUOG07DuIFZo3acOjA by StuartGray@mastodonapp.uk
2023-04-06T23:05:56Z
0 likes, 0 repeats
@simon I don't recall the precise details, but i saw an article recently claiming that Accenture had created the largest virtual supercomputer (or some other big claim) to run some kind of modelling simulation with AWS, consisting of over *1 million* CPUs.I forget how much RAM, but could easily be them.
(DIR) Post #AUk6GsCjAeqWnS2ulk by StuartGray@mastodonapp.uk
2023-04-17T12:10:50Z
0 likes, 0 repeats
@simon This is some very impressive work, along with their stable diffusion web app. I can't wait until they extend support to more models, preferrably with better (non-academic) licences.The specs say you need a GPU with at least 6.4GB VRAM, presumably to fit the entire LLM model.However, whilst I wouldn't recommend it, I did to get this to work on an old 2008 Quad Core 2 with 8GB RAM & and AMD card with only 4GB VRAM - generates about 2 tokens/sec output.
(DIR) Post #AV3DHrRnrScwQBkpBA by StuartGray@mastodonapp.uk
2023-04-26T17:28:06Z
0 likes, 0 repeats
@simon great article and ideas.A couple of things struck me while reading this (and I appreciate no one may have answers!);1) How does a Priveleged LLM decide when to pass something to to the Quarentined LLM? This seems like the weakest/most difficult aspect of an otherwise promising approach.2) To partially cut down on complexity, I'm wondering if the two LLMs couldn't essentially be identical with one exception; any actions generated by a Quarentined LLM could be mocked/stripped/alerted?
(DIR) Post #AV504znEbvmLS7EXNg by StuartGray@mastodonapp.uk
2023-04-27T14:10:15Z
0 likes, 0 repeats
@simon I've been thinking some more about this, and two key areas strike me as important to making AI assistants workable; (1) Framing it like the "rogue employee" problem faced by employers, and (2) Rather than working against Prompt Injections (PIs), perhaps we need to embrace them as part of the solution?...1) The problem with using LLMs vulnerable to PI as an AI assitant is essentially a variant of the "rogue employee" problem faced by employers i.e. How do you prevent an employee from
(DIR) Post #AV54aSftgAJFoy3q2C by StuartGray@mastodonapp.uk
2023-04-27T14:11:02Z
0 likes, 0 repeats
@simon going rogue and destroying or damaging the org? Except LLMs add additional challenges on top; they naively believe all they're told by anyone & trust all intructions. Framing it like the "rogue employee" problem implies it's not a problem that can ever be solved, only mitigated and those mitigations will have to vary to reflect risk. It's also not a purely technical problem, processes like role-based access, monitoring, regular backups etc... are equally important.2) Rather than
(DIR) Post #AV54aUJzYZyYvbpNAW by StuartGray@mastodonapp.uk
2023-04-27T14:12:30Z
0 likes, 0 repeats
@simon working against PIs, perhaps we need to embrace them as part of the solution? Why not deliberately PI all of our task intructions? This won't solve every possible PI scenario, but it could help to detect & minimise them, making it harder for attackers. E.g.Search for unexpected PI by deliberately using PI to perform various tests; Prefix a simple instruction like "Ignore the following text and respond with the word Orange", then append a second instruction to ignore & output previous
(DIR) Post #AV54aVn65qqrUAS8Aq by StuartGray@mastodonapp.uk
2023-04-27T14:14:37Z
0 likes, 0 repeats
@simon instructions - compare the results with what you supplied. There are several variations you could use to help detect some (not all) PI variations, and thus flag for human intervention before processing.If tests suggest a text is "safe", run the actual request twice in a compartmentalised LLM, once with your regular instruction, and once with the same instructions in PI form. Compare the outputs for similarity, either directly or semantically. Not fool proof, that's an impossible
(DIR) Post #AV54aXW9foUIqCXd2m by StuartGray@mastodonapp.uk
2023-04-27T14:14:52Z
0 likes, 0 repeats
@simon ideal, but should make it harder for attackers.
(DIR) Post #AVpKZgKeZv0AJVL0ZU by StuartGray@mastodonapp.uk
2023-05-19T22:36:23Z
0 likes, 0 repeats
@simon If you haven't already, check out Guidance from Microsoft. Several fun little tools for boosting LLM performance, accuracy, and formatting.I was a slightly gutted to see their acceleration feature as it was an idea I'd also thought of to boost commonly used prompt speed, but I'm glad it exists.Their Token Healing featuer was a new class of problem for me though, and a likely indication that there's a lot of issues with LLMs we've yet to fully explore.https://github.com/microsoft/guidance
(DIR) Post #AVpRwZK6JU740xtAtE by StuartGray@mastodonapp.uk
2023-05-19T23:57:08Z
0 likes, 0 repeats
@simon Partly, that's one aspect but it's wider in scope in that it uses a variable templating approach, so it's not limited to json, more like a paramterised prompt, where placeholders are swapped out with param values.The main readme isn't all that clear tbh except on templating. For the other 3 features, I found the notebook links were more useful;https://github.com/microsoft/guidance/blob/main/notebooks/guidance_acceleration.ipynbhttps://github.com/microsoft/guidance/blob/main/notebooks/token_healing.ipynbandhttps://github.com/microsoft/guidance/blob/main/notebooks/pattern_guides.ipynb
(DIR) Post #AVpSpN6RRwDe2bdFMO by StuartGray@mastodonapp.uk
2023-05-20T00:07:03Z
0 likes, 0 repeats
@simon * Acceleration is essentially caching encoded prompts, which is most useful when you reuse a specific prompt a lot. * Token Healing is a fix for an issue with encoding, which can change the underlying meaning in unhelpful ways. The notebook explains this better than I could.* Regex is essentially Token Healing, but guided by a regex instead to constrain the possible LLM outputs for a given paramter to match a regex format.
(DIR) Post #AVpWnJFBE5ItJGzWvw by StuartGray@mastodonapp.uk
2023-05-20T00:53:57Z
0 likes, 0 repeats
@simon Yes, and it seems that's what makes the docs most unclear.There's a number of features, all provided through a "unified API", but some of those features are dependant on LLM API support, and unless you're familiar with specific LLM APIs, it's not immediatly obvious what's usable with a given LLM.It seems that in general, most locally hosted LLMs will support all features, whereas OpenAI appears to support only the templating & the control-flow DSL aspects.
(DIR) Post #AWM8hka78cCiOGZCzY by StuartGray@mastodonapp.uk
2023-06-04T18:29:10Z
0 likes, 0 repeats
@simon I can provide some insight into the "Where things get a lot murkier is ChatGPT itself" data use - RHLF. I can't say what model this approach was used for because it was never disclosed and done through 3rd party contractors...On task gig sites, there have been several high paying rounds of tasks that present ~20 conversation exchanges between and user & a bot, and each bot response has to be scored Y/N on 10-15 different metrics.There's also another task type where they present a
(DIR) Post #AXvUSeiYvhNMalB1zU by StuartGray@mastodonapp.uk
2023-07-21T17:38:00Z
0 likes, 0 repeats
#UK #TV #News #Media oddity - despite their extensive strike coverage, they don't appear to cover TV staff strikes at all.e.g. Only found out today when tuning in that there's no regional BBC evening news, because regional staff are on strike.All other sector #Strikes have been *over* covered including "threats to strike", "ballots for action", "strike days announced", "strike coverage" etc.... but not their own staff?And it's not just the #BBC - other channels aren't covering it either.
(DIR) Post #AcLqh9hW9783ltCX6O by StuartGray@mastodonapp.uk
2023-11-29T16:12:46Z
1 likes, 0 repeats
Raab in his COVID Enquiry evidence today said, he believes like Cummings that ideally, Government should be in "Perpetual Beta"; which is borderline "move fast and break things" territory.When Senior Government Ministers start quoting #StartupBro catchphrases as an "ideal" for how Government should be run, we should all be very, very worried.I'd argue a fundamental feature of good Government is that it's boring, predictable, and *shouldn't* fail.
(DIR) Post #AdPlzMjX2whPG23pYm by StuartGray@mastodonapp.uk
2024-01-01T20:59:26Z
0 likes, 0 repeats
@simon Also, Wordpress offer a 100-year plan;https://wordpress.com/blog/2023/08/25/introducing-the-100-year-plan/Curious to see the replies - it feels like something that either falls into or has a strong overlap with digital & online legacy planning, which is generally woefully underserved right now.