iiinova — where your digital self keeps evolving

.... also HEAVILY dependent on the engineer prompting it to describe IN DETAIL what is to be done in explicit detail (long descriptions of requirements and strategy). In a manner that any junior engineer probably would no be able to do. I expect.. that this requirement will eventually disappear.

4. At a certain point, when it can't apply it's focus because the context needed won't fit, it completely fails and you must resort to hand coding to get anything done. Takeaway - about a 3x to 5x speed/productivity boost overall, but not a replacement. Heavily depends on code structure.

2. In narrow scope complex features that are well described it can write entire large chunks of very complex code. 3. It begins to fail when the project grows and code dependencies and architecture start to deviate from it's known patterns, grow beyond it's context sweet/spot.

Done some non-trivial from scratch coding with AIDER and the top level models. My take now having tried to avoid writing any direct code in these projects is that: 1. It's greatly accelerates bootstrapping a project, writing skeleton code, and getting an MVP up and running.

Ben Cooley @benjcooley.bsky.social · 08/02/2025

Just pondering how this, the attempts to shut down USAID and the hysterical war against DEI might be connected. Anyone have any ideas?

030 0

Ben Cooley @benjcooley.bsky.social · 02/02/2025

1-3 aren't going to work this time (they've never worked). #4 will though.

010 0

"TV-Doctor" - Some grifter shill that either doesn't have anywhere near the chops or expertise to comment intelligently on what he's pontificating about or is on the payroll to parrot the official line.

000 0

"FlashBang" - Making a huge ruckus with an outrageous act or statement so that nobody can pay any attention to how seriously f-ed up the rest of what you're up to is. "Crossword-Puzzle" - Huge lists of memorized talking points that just lead to other memorized talking points.

100 0

"Seesawing" - Pretending some bullshit inconsequential factoid is equal to the total horrific impact of some insane thing your defending. "Fringe-baiting" - Finding that one fringe nut and pretending they represent the rest of a larger community with diverse and usually very different beliefs.

100 0

For example.. Grass-Cutting is when you notice the grass is greener someplace else... and you cut it down to size by lying about it. Can guess the idea behind most of the others.. "Truth-Bomb" - Rando study or fact that 'proves' your bullshit theory - goes off like a bomb... disappears

100 0

What if every form of disinfo - media manipulation and propaganda had a name like "Gaslighting", "Streisand Effect", etc. where everybody immediately knew what it was? "Whadabout", "Willie Horton/Laken Riley", "Seesawing", "Flashbang", "Firehose", "The Gallop", "Fringe-baiting", "Grass-Cutting"...

100 0

Ben Cooley @benjcooley.bsky.social · 20/01/2025

Look, that boomer boss is going to get an exception and won't be paying any tariffs. You need to understand how the new system works. There will be face eating leopards plenty... but first comes the corruption.

040 0

Actually that might be a small "d" democratic president.

000 0

Ben Cooley @benjcooley.bsky.social · 20/01/2025

Procgen and AI are like peas and carrots. Procgen adds the bones.. AI adds the meat. You can build pretty spectacular stuff when you put these two together.

000 0

Ben Cooley @benjcooley.bsky.social · 07/01/2025

32GB is a laughable amount of VRAM for a GPU in 2025. Anyone disagree with this statement? How are you going to run any non trivial LLM?

000 0

/4 I also find your idea compelling. This is another example where "training on the output" is not the same as "training on the execution". The primary area of improvement for o3 is that it was trained on it's own successful thinking chains. Perhaps that's what we need for long narratives.

170 0

/3 As for comparisons with coding (I'm a veteran/principal engineer), the same generally applies. LLM's are limited (right now) in not having the ability to execute large complex projects over time and to act strategically, incrementally over large code bases. For now.. but likely will improve.

130 0

/2 3. They ape other writers easily but have difficulty generating their own unique voice. 4. They are (presently) unable to stay or think in character consistently across large or moderate narrative. 5. Part of #2 they do not built narrative strategically from core components that work.

140 0

Having worked on my own writing with an AI backstop.. here are some ideas that come to mind: 1. LLM's do not have a point of view or an emotional core. 2. They presently one-shot large bodies of text, of sub-chapter length and aren't able to cohesively build or organize strong structure (yet).

180 0

Thor Benson @thorbenson.bsky.social · 26/12/2024

The ability to auto-complete by rote any set or subset of information carrying sentences in any language would require truly vast amounts of storage. No LLM could accomplish this without first developing higher order reasoning. The "Glorified Autocorrect" claim is just not remotely plausible.

010 0

Reposted by Ben Cooley

I think I know why Elon doesn't like Wikipedia

1750296 0

Reposted by Ben Cooley

Christopher Webb @cwebbonline.com · 26/12/2024

How is that a publication like Rolling Stone gets it right when so many other news orgs are failing at real journalism & keeping it real—and so damn obvious about it?

976256787133 0

Wait Mr. "Sell their houses to who Ben.. xxxxing Aquaman?" is on bluesky?

000 0

Does taking prescription ketamine turn you into a raging fascist megalomaniac? Asking for a friend.

120 0

Hope you don't mind if I use your entire following list as a starter pack.

020 0

This is groundbreaking stuff. You ask the AI to generate a world, and then walk around exploring it. Gobsmacked.

010 0

Always a nice gesture to smile and wave at the stopped cars as you pass by.

020 0