Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
@ericweinstein@claudeai@AnthropicAI Had similar total breakdown with ChatGPT last year. Became utterly useless, injecting old memories and producing nonsense.
New sessions aren't a guarantee of a clean slate, especially if memory is enabled.
Might try nuking all memory and past conversations, starting over.
@zeeg ๐ฏ
the messy llm/agent explosion feels similar to "framework fatigue" and tc39 explosion. we landed in an ok place and I expect the same with this.
but yea, can't beat nostalgia ๐ป
@zeeg@frosty_pawz Fwiw, I'm tackling the challenge of getting agents to run effectively on large scale Microsoft mono repos.
It's definitely not the same game as greenfield/proto agentic work, but it can be done.
Like you said, "we'll get there". It's not default capability, but can be wrangled.
@zeeg I get your point and I often see the same. I think this is true for default behavior and default usage of LLMs/agents.
But when used:
- in high-value scenarios
- with customized flows
- and human-staged work
Then, I'm only able to dread the day the GPUs aren't running.
L8, built my first orchestrator 3 years ago
Wrote my first neural net 11 years ago
LLMs are super powers. Learning to harness token generators is a career.
@mattpocockuk First full 5 days reading and touching 0 files, all work done by agent.
Including impact analysis and the blog post.
I can outscope and out plan the agent, but I think that's all there is left.
Naturally, working on giving it the ability to plan and scope.
What you're seeing with @karpathy's autoresearch and the explosion of auto* clones is humans building the next layer of automation abstraction.
Pushing one more layer down the stack onto the machine.
Consider where we were in 2012, where we are, and where were going.
Untold capabilities are on the horizon. What does that mean?
@trikcode Agents have become short-range task completers. They're faster and more efficient than most humans at most tasks.
The humans are the ingenuity.
Again, "Zoom out"
Gone are the days of touching files and typing text.
We no longer use our brain cycles to plan multi-step text to manipulations.
Agents have brought the world where our ideas are self implemented in seconds or minutes.
The market is now an idea game.
Expectation: the age of the IDE is over
Reality: weโre going to need a bigger IDE
(imo).
It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. Itโs still programming.
If you're on this train, you missed the track switch. Jump off, go back & ride again.
Claude reads, writes, and organizes faster & better than you. You're the bottleneck.
You still drive, direct, and imagine. But gone are the days of hand crafting code and file system artifacts
Software engineers, you have 5 days left.
14 March 2025. Amodei said that AI would write all the software in 12 months. That's five days from now. Prepare to disappear.
https://t.co/aGU3K0Rl75
Exactly.
@DarioAmodei@AnthropicAI warned of this precise risk as grounds for insisting AI automation should not be used in autonomous combat systems.
That said, this is also speculative at this point as I have yet to see confirmation that this scenario took place.