Michael Hartle

@glyph@mastodon.social I think that agentic development will affect software development, from a systems-thinking perspective (see Donella Meadows "Thinking in Systems"). In a software project with "guardrails" (architecture / unit / integration / e2e tests, formatting rules, sane build scripts, documentation, ...), letting a set of LLMs run wild iterating towards a testable / measurable goal, yielding a MR that is reviewed by senior engineers is akin to computationally brute-forcing software development. 1/n

@glyph@mastodon.social Companies swallowing the AI promise hook, line and sinker enter a strategic dependence that will be hard to quit. This hands a big leverage to LLM providers in need to generate an ROI. Compare the Web how it started to its current surveillance-focused form, especially how products and services tend to follow the path of @pluralistic@mamot.fr s Enshittification. 4/n

@glyph@mastodon.social Now take a look at companies having transformed themselves, deeply integrating their knowledge bases, their engineering, strategic decision and operational processes with LLM providers. Atrophying their organizational capacity and muscles to decide and act independently, becoming "brains in jars". Connected to LLM providers specialized in learning and mimicking, with every incentive to increase their shareholder value, having proven that they don't care about copyright or stuff. 5/n

@glyph@mastodon.social This is the point where the cost of brute-forcing software development will be no longer irrelevant - for developing software, for the company and its competitors, for people, for the environment. For me, the question is how to navigate well past this phase, and how not to end up in the hands of extractive monopolies. 6/n

@glyph@mastodon.social Some companies transform themselves. Cory Doctorow pointed out in his Guardian article that AI salesmen just need to be convincing that devs can be replaced by LLMs, with juniors becoming optional and seniors the accountability sink ("human in the loop"). But the outcome here is different. A company that does not invest into its juniors becoming seniors, and being heavily dependent on its LLM provider. If enough companies go that route, it becomes relevant systemically. 3/n

@glyph@mastodon.social Given a harness exists that enables a tight feedback loop between probabilistic LLM output and the harsh reality of compilation errors, hexagonal architecture rule violations or integration failures, brute-forcing is a possibility if and only if the cost is sufficiently low so it does not matter. From my point of view, a lot is going on. Massive datacenter investments expect an ROI. The pressure to get much more tokens used at a much higher price is huge. 2/n

@ilumium@eupolicy.social Well, belief is a big thing in both domains.

@jwildeboer@social.wildeboer.net This ties nicely into understanding AI prompting as a Skinner box. Like slot machines in Las Vegas, loot boxes in online games or booster packs in trading card games, success & gratification is probably just one more try away...

@glyph@mastodon.social I hope they can fight for measures to prevent vendor lock-in, and be able to use the quality / cost monitoring to get an objective early warning before the extraction phase of LLM dependence hits with full force. 8/8

@johncarlosbaez Is this a compensation for them using your work illegally, barring them from further use and keeping you in control, or a compensation that "retroactively licenses" usage of your work in Anthropics current and future LLMs?

@glyph@mastodon.social Meanwhile, even in companies going all-in on "AI", devs are needed - at least to build and refine harnesses that allow for brute-forcing in the first place, to review the output and guide where LLMs go wrong and to.fix the most strange, unexplained incidents. For navigating through that phase, I hope they will get a chance to make the quality and cost of LLM usage objectively measurable and transparent to their management. 7/n

@johncarlosbaez So this is a settlement mechanism requiring you to do stuff the publisher should do, share compensation with the publisher, would prevent you from joining future lawsuits and becoming binding by default on inaction? Given how some people and companies got sued into oblivion on digital piracy for a lot less, this sounds like quite a bargain for Anthropic.

@ewolff@mastodon.social @shanselman@hachyderm.io Unfamiliar with the environment, yet missing an account with proper roles, no test systems, forced to meddle in production with high stakes, unclear goals, documentation not at hand, bad internet connectivity, high-pressure workplace, so a pretty bad onboarding, any crash gets you management attention, no way to meaningfully influence management to improve your situation. Not unheard-of initial IT project conditions for a multi-year project compressed into <15 hours.

Also available as digital edition in EPUB and PDF format at https://www.informit.com/store/architecture-for-flow-adaptive-systems-with-domain-9780137393039

@danimo@chaos.social Argh, danke für den Hinweis, das war mir nicht bewusst - damit endet diese Geschäftsbeziehung.

@cwebber@social.coop Well, people went to StackOverflow with a question and looked forward to answers based on the experience of others. While one can still ask an LLM and give a rubber-duck training session for its provider, I still fail to see the influx of answers based on experience.

@Extelec@mstdn.social go west

@MissingThePt@mastodon.social The originally intended destination of Twitter at the time of its web disappearence is still unknown. The maximum depth of rock bottom in this region has not yet been established, either.

@ratgrrl @jeffjarvis Somehow a strange position of the quote; a #LLM is a tool, just doing what it does, and at the same time, the quote falls back to stating this behaviour is not its fault, as if it was an entity capable of actual reflection, or even taking responsibility for its fault. Responsibility for all this lies with who did its development, release and marketing.

@mhoye@cosocial.ca You can make others aware about something, and explain your reasoning; everything else is a bonus. I have made (and am constantly re-making) the experience that people sometimes need time and 7 different ways of explanation to see what is an obvious given for you. Sometimes they will never share your perspective at all.

@janboehm@edi.social Er sollte so konsequent sein und das Werbefotografie-Model werden, was er immer sein wollte - dann könnte ich ihn konsequent mit der Werbung ignorieren.

@carnage4life@mas.to What do we say? Not today (on "my" codebase)!

@Philsturgeon@mastodon.green You mean, after having invested into AI infrastructure due to FOMO and not being anything close to covering these costs in the foreseeable future?