theptip

16 ms·

361.
by theptip 10mo ago
It’s true - but there are “good code” solutions to this already. For example, BDD / Acceptance Tests can be used to write human-readable specs. IMO it’s quite boilerplate-y to set this up pre-LLM but probably the ROI is favorable now.
362.
by theptip 10mo ago
Ad noob here; how would you test this hypothesis?
363.
by theptip 10mo ago
It’s a good point, if they are actually struggling then a fire sale is justified. The counterpoint would be that 3 months ago they raised $750m at a $6.9B valuation. Unless they are burning through that cash so fast that they need to start
364.
by theptip 10mo ago
> The cost of that suit will be more than the value of the gains and damages awarded. In many cases this is so, but here we are talking about tens of billions in value. Even a few percent of value won is worth lawyering up to the hilt fo
365.
by theptip 10mo ago
Don’t the founders (and the board) still have fiduciary duty to the common holders? You can’t stop the founders from leaving, but selling the crown jewel IP in a transaction that doesn’t benefit the shareholders seems a stretch.
366.
by theptip 10mo ago
It’s also possible that both are correct, and the deal is actually illegal. It’s pretty common for deals to push close to the line to extract maximum value for one set of parties, and sometimes this is misjudged. I guess we just need to wai
367.
by theptip 10mo ago
https://tools.simonwillison.net/
368.
by theptip 10mo ago
Yes, check again - to be blunt, any opinions (at least tactical on how well feature X works) formed 6 months ago are not really relevant to the conversation today given how fast this is all moving. Opus 4.5 in Claude Code is a massive jump
369.
by theptip 10mo ago
Or it’s conclusive of an even broader trend!
370.
by theptip 10mo ago
A really interesting point that keeps coming up in discussions about LLMs is “what trade-offs need to be re-evaluated” > I also believe that observability is up for grabs again. We now have both the need and opportunity to take advantage
371.
by theptip 10mo ago
Doubling time has been 7 months for a while, so you should expect 8h not 1 week next year.
372.
by theptip 10mo ago
You don’t even need to go this far. Just make deletion a right and clear consent a requirement like GDPR did. That’ll kill all these systems that depend on collecting information about people without their knowledge. (Same goes for the cred
373.
by theptip 10mo ago
Disagree. The journals are now acting like a paid certification. If they admit any old slop, who would pay to submit their papers? The service they are providing is peer review and applying a reputable quality bar to submissions. Think of i
374.
by theptip 11mo ago
> very few people can correctly articulate requirements The observation from Lean is that the faster you can build a prototype, the faster you can validate the real/unspoken/unclear requirements. This applies for backends too.
375.
by theptip 11mo ago
The point with clippy is just that the AGI’s goals might be completely alien to you. But for context it was first coined in the early ‘10s (if not earlier)when LLMs were not invented and RL looked like the way forward. If you wire up RL to
376.
by theptip 11mo ago
Exactly. Building a model that truly understands humans, and their intentions, and generally acts with, if not compassion then professionalism - is the Easy Problem of Alignment. Starting points: https://www.lesswrong.com/po
377.
by theptip 11mo ago
I’ve seen this result. I wonder if it’s because LLMs are (grok notwithstanding) deliberately middle-of-the-road in their stances, and accurately and patiently report the facts? In which case a hypothetical liar LLM would not be as persuasi
378.
by theptip 11mo ago
The increase in myocarditis from the vaccine is well-documented. (And very small.) COVID causes myocarditis too (even for young people unlikely to die from COVID itself), at much higher rates. So you only need a 20% chance of contracting CO
379.
by theptip 11mo ago
It takes guts to say “this 1 month feature would be done in a couple days by a competent competitor using modern technology and techniques”, and the legendary “I reimplemented it in <framework> over the weekend” is often not well rece
380.
by theptip 11mo ago
What are the trade-offs you’ve seen between the two? Always appreciate this sort of experience report on HN!
381.
by theptip 11mo ago
> Cursor had won the developer market from the previous winner copilot It’s a fair point, but the counter-point is that back then these tools were ide plugins you could code up in a weekend. Ie closer to a consumer app. Now Claude Code i
382.
by theptip 11mo ago
Retail is a big deal these days. Used to be sub 10%, now it’s in the 30-40% of daily volume range IIUC. You can easily look up the numbers you are asking for, the TLDR is that the volume in most stocks is high enough that you can’t manipul
383.
by theptip 11mo ago
It’s not gonna happen any time soon. The model is fine-tuned on traces generated by the scaffolding (eg dependent on what tool calls are available), and the scaffolding is co-developed with the strengths/weaknesses of the specific mode
384.
by theptip 11mo ago
But TRT suppresses endogenous production further, so if you go off it you’re worse than when you started.
385.
by theptip 11mo ago
My current idea for this is to have AI administer the 1:1 oral exam. I’m quite confident this would work through grade school at least. For exams you’d need a proctored environment of some sort, say a row of conference booths so students ca
386.
by theptip 11mo ago
Yeah, these AI detector tools are bad. (They also punish ESL as this tends to result in simpler language patterns that parse as LLM.) The great thing about AI is that with a bit of imagination it can be used to amplify teachers too. In this
387.
by theptip 11mo ago
IMO for “boring software” you usually want to be on the oldest supported main/minor version, keeping an eye on the newest point version. That will have all the security patches. But you don't need to take every bug fix blindly.
388.
by theptip 11mo ago
The general pattern seems to be that LLM+scaffolding performs better than LLM. In 6 months time a new model will incorporate 80% of your scaffolding, but also will enable new capabilities with a new layer of scaffolding. I suspect the model
389.
by theptip 11mo ago
What’s your opinion on any of the plethora of unicorns in domain-specific AI, like Harvey? ($100m ARR from what I could find on a cursory search) https://www.forbes.com/sites/iainmartin/2025/10/29/le
390.
by theptip 11mo ago
Yeah, I believe that lots of startups don’t have evals either, but as soon as you get paying customers you’re gonna need something to prevent accidentally regressing as you tune your scaffolding, swap in newer models, etc. This is a big cha

More ›