gordonhart

8 ms·

91.
by gordonhart 8mo ago
Honestly curious, have you seen agents succeed at this sort of long-trajectory wide breadth task, or is it theoretical? Because I haven't seen them come close (and not for lack of trying)
92.
by gordonhart 8mo ago
Codex with their flagship model (currently GPT-5.3-Codex) is my daily driver. I still end up doing a lot of steering!
93.
by gordonhart 8mo ago
Curious to hear if you've seen this work with 100k+ LoC codebases (i.e. what you could expect at a job). I've had some good experiences with high autonomy agents in smaller codebases and simpler systems but the coherency starts to
94.
by gordonhart 8mo ago
Here's an example ticket that I'll probably work on next week: Live stream validation results as they come in The body doesn't give much other than the high-level motivation from the person who filed the ticket. In o
95.
by gordonhart 8mo ago
Whenever I get worried about this I comb through our ticket tracker and see that ~0% of them can be implemented by AI as it exists today. Once somebody cracks the memory problem and ships an agent that progressively understands the business
96.
by gordonhart 8mo ago
I kind of agree with GP more than the author here. OpenAI got these people hooked and pulling the plug is potentially more harmful than letting them continue to chat with it until they move on naturally (assuming that they eventually will)
97.
by gordonhart 8mo ago
Releasing the weights is an easy and low-cost way for OpenAI to fix this problem.
98.
by gordonhart 8mo ago
That's... a strategy. Matter of time before an AI companion company succeeds with this by finetuning one of the open-source offerings. Cynically I'm sure there are at least a few VC backed startups already trying this
99.
by gordonhart 8mo ago
Anecdotally, 4o's sycophancy was higher than any other model I've used. It was aggressively "chat-tuned" to say what it thought the user wanted to hear. The latest crop of frontier models from OpenAI and others seems to
100.
by gordonhart 8mo ago
Yeah, given that frontier model training has shrunk down to a handful of labs it seems like a very solvable problem to just build the stack directly without CUDA. LLMs are mechanically simple and these labs have access to as much engineerin
101.
by gordonhart 8mo ago
The main coding agent failure modes I've seen: - Proliferation of utils/helpers when there are already ones defined in the codebase. Particularly a problem for larger codebases - Tests with bad mocks and bail-outs due to missing t
102.
by gordonhart 8mo ago
For better or worse, a lot of people seem to disagree with this, and believe that humans reading code is only necessary at the margins, similarly to debugging compiler outputs. Personally I don't believe we're there yet (and may n
103.
by gordonhart 8mo ago
Mass panic? Think of how wildly it would be misrepresented in the media and how disruptive it would be to all air travel in the country. People aren't rational actors and the most sensationalized headline is what ends up spreading
104.
by gordonhart 8mo ago
Great project, thanks for building and sharing!
105.
by gordonhart 8mo ago
The difference I see is that, after "get to orbit", the goalposts for SpaceX are things that have never been done before, whereas for LLMs the goalposts are all things that skilled humans have been able to do for decades.
106.
by gordonhart 8mo ago
He responds to a thoughtful and detailed 600-word comment from a maintainer with a dismissive "Here's the AI-written copyright analysis..." + thousands of words of slop. The effort asymmetry is what's rude. The maintaine
107.
by gordonhart 8mo ago
Personally I’m less bullish on oil as the metaphor given how much of the modern world is underpinned by cheap and ubiquitous oil. If oil disappeared tomorrow, society would collapse — if tobacco disappeared tomorrow, it would make some subs
108.
by gordonhart 9mo ago
On a scale that would make big tobacco blush.
109.
by gordonhart 9mo ago
Populations grow logistically as they hit carrying capacity, it just looks exponential when it’s far from K
110.
by gordonhart 9mo ago
Clawdbot/Moltbot/OpenClaw is so far from figuring out the “trust” element for agents that it’s baffling the OP even chose to bring it up in his argument
111.
by gordonhart 9mo ago
Same, it's nice to use a no-BS CDN for personal projects (e.g. https://atlasof.space/ ). Their pricing is good and I actually appreciate that they have no free tier so that there's no "oh shit" moment whe
112.
by gordonhart 9mo ago
By the time that happens, Opus 5 and GPT-5.5 will be out. At that point will a GPT-5.2 tier open-weights model feel "good enough"? Based on my experience with frontier models, once you get a taste of the latest and greatest it&#x2
113.
by gordonhart 9mo ago
I'm not really interested in getting in a political mudslinging contest; you asked what's to stop a future president from doing so and I gave a practical answer.
114.
by gordonhart 9mo ago
SpaceX launches about 60% of National Security Space Launch payloads. The only other active launcher cleared for these missions is ULA's Vulcan Centaur, which has launched 3 times total, ~once every 8 months.
115.
by gordonhart 9mo ago
That’s been my reaction to every Slate Star Codex/Astral Codex Ten piece I’ve read.
116.
by gordonhart 9mo ago
Shocked to see SpaceX buy the datacenter in space meme. Where does the power come from? Where does the heat go? Why add (high) launch costs to your buildout capex? Why add radiation as another risk factor to your already-unreliable GPUs? Am
117.
by gordonhart 9mo ago
Hinges on whether this new high level -> low level transformation becomes reliable enough to build watertight abstractions on top of it. If AI code becomes good enough that you don't have to worry about the low-level representation
118.
by gordonhart 9mo ago
As far as I can tell, it’s a viral marketing scheme with a shitcoin attached to it. Hoping 2026 isn’t going to be an AI repeat of 2021’s NFTs…
119.
by gordonhart 9mo ago
If you’re like me and get inundated with spam texts from politicians begging for money, you can set up an app to filter texts from unknown contacts matching content patterns. I recently set up SpamHound but there are many options, and confi
120.
by gordonhart 9mo ago
GP was making a joke, but Anthropic could implement this if they wanted to. Not a bad metric actually if you can measure it cheaply enough.

More ›