Aldo Cortesi

Announcing spacecurve, a space-filling curve library with a web + native interactive playground.

@alexdong Tuesday? I have many thoughts.

@cortesi looking forward to hearing all about it.

Finally installed OpenClaw, and I have some thoughts. youtube.com/shorts/TmfVDYu…

@bdmorgan Can you clarify whether driving Gemini CLI through ACP for personal agentic setups is within acceptable use with this change?

Gemini CLI users - please see this important announcement here: github.com/google-gemini/… Quick summary: ( 1 ) Users using the product for free ("Code Assist for Individuals") will lose access to Pro models on March 25. ( 2 ) We will be increasingly prioritizing traffic based on subscription type and account standing.

@JackWoth98 Can you please clarify whether driving Gemini CLI through ACP for personal agentic systems is permitted?

An important update for Gemini CLI users 👇 To be transparent we have had to make some tough decisions to keep up with demand and provide a stable service for higher trust/priority subscriptions. Most notably the free tier of Gemini CLI, which in a weeks time will lose access to Gemini Pro models. Definitely worth reading the full details, if folks have questions, please ask them and I will do my best to respond 🙏

@NinaPanickssery Cheating, or just efficient? The great question of our time, I feel.

@cortesi this is cheating though but yeah correct

I wonder if people can guess where this photo was taken

Of course, important not to lose sight of how completely nuts this is. I mean this is Terry Tao vibe-formalizing a Lean proof with Claude. The world is going to get very strange, much sooner than most people expect.

So charming. The greatest mathematician of our age getting anxious when Claude spends 30 seconds thinking. "I'm sure it's burning a lot of tokens...". Meanwhile Claude has spent 45 minutes futzing with CSS for my side project, and is not yet done. youtube.com/watch?v=JHEO7c…

Facing the fact that many clever, tasteful technological choices I made are obsolete. The tech stack for maximum agentic velocity is completely different from what I'd use if it was just me.

The world is about to get very, very weird.

would be ideal, but we've both been around long enough to see 1000 good-hearted security and privacy projects break themselves on identity validation. i'm pessimistic about solving it. if a system like this is local, private and curated by a human and their agent with private selection criteria, i'm not sure how much of a practical problem bots will be. if it proves problem in practice, one could still play with transitive trust scores and discovery through mutuals. anyway, it's so easy to experiment now that a gung-ho dirty solution first with rapid iteration is probably still the right approach.

the problem is that bots can easily masquerade as seemingly useful for a while, get followed, and then turn spammy, or just slightly spammy enough, etc. sometimes 50% of user activity on delicious was inauthentic, and that was 20 years ago. just detecting it will be very hard going forward. i want a way to tie accounts to some specific identity (even if that identity is not obvious to other users of the system)

15 years ago HN was annoying but informative - now you can read hundreds of comments on a major announcement and every single one is wrong or spiteful or both. Is constructive technical conversation just dead now? Has it moved somewhere I don't know about?

maybe i'm being too aspie about this, but i wonder if the identity issue can treated with a loose touch. i'm thinking of a sytem where an identity is roughly a source (of links, comments, quotes over all and any platform, including email), and you calculate a quality metric, of whatever kind you prefer, over that source. if it's high it's boosted if it's low it's deboosted or removed. so i add "joshu" here on twitter, my agent goes out and finds your social presence approximately everywhere you publish, assembles you into a source. your output is consumed through an agent-driven web browser, so there's no ability for platforms to silo. if i respond, the agent responds natively on the appropriate platform on my behalf. if there's an impostor or something squatting joshu somewhere else we can detect it and clean house. i think this way one could pull of a heist that lifts your social or informational graph out of the silos while they burn to the ground.

@cortesi yeah. we should have solved identity before now, but it’s gonna become a festering wound.

@joshu right. and an increasing number of the comments are AI anyway. i don't think any open forum will survive intact - we're going to need protected enclaves with curated person-to-person trust relationships that sidestep silos. now's the time to try to alter the deal. exciting times.

@cortesi definitely. The criticism dynamic seems to be an engineering/tech dynamic. Half the the comments on hn are just rants that are only vaguely triggered by the content of the article. Google did something -> rant about google privacy shit

The social dynamic is interesting. Criticism is social currency, but that currency can only be checked in with the spectating crowd. I've always felt that smaller networks where everyone's invested were more valuable, but they had to co-exist over the top of the baying crowd because of network effects. It might be possible to take a different approach now with AI.

@cortesi criticism is social currency

@thsottiaux Key binding configuration for the codex native macos app. I'd like to use it, I really would, but I just can't do it without being able to configure key bindings the way I want. In general more expert features for the app: tabs, multiple windows, key bindings.

With GPT-5.4 out. What should Codex ship or improve next?

I've been thinking along similar lines. I want to to build a tool that lets you collect interesting people and then collate their activity whatever platform they're publishing on, whether it be X, Mastodon, blogs, Facebook, email newsletters, whatever. A platform-agnostic omini-follow, with a layer of scoring/filtering/collation/presentation built on top to surface what you are interested in from trusted sources, detect signal drops, manage your sources smartly. AI is going to break down all of our social silos, but if we can use it to build out personal networks the world might even be better for it. We're all desperate to escape the stranglehold of social platform network effects, and for the first time I actually feel like it might be possible.

@cortesi I largely agree with you. But my browser extension Hacker Smacker (which was on HN last week) makes it a lot easier to spot the good commenters when they show up again in later discussions. They’re still there, just among a larger sea of noise.

Evaluating GPT 5.4 this morning, and it's great - another increment that seems more and more like a step change the harder I look. OpenAI is pulling ahead strongly now, hotly pursued by Anthropic. Google is astonishingly far behind - what are they doing??

@Teknium This is especially frustrating with Deep Think. I've exhausted my daily Deep Think queries multiple times trying to figure out which part of my prompt or files causes it to just barf and do nothing.

Love it when Gemini just discards my prompt + any images/attachments and then doesnt show any error or otherwise that it broke or why

@nisten I just haven't found Deep Research variants to be useful for this kind of query - they're more aimed at discursive/narrative overviews and lit reviews. Did you try ChatGPT Pro?

New one shot benchmark for you to check how every single Deep Research report will make mistakes in numbers, wont obey to you asking for evidence for everything single entry on the table even though they're all making 300+ queries each. Prompt: Make me a comparison table for Apple chips M1-M5 max with this exact lenght and these exact columns in this EXACT format and check every single number you put in there: +---------------------------------------------------------------------+ | COMPARISON OF ALL APPLE M CHIPS M1-M5 MAX | +----------+-----------+--------------+--------+-----------+----------+ | SoC | Max CPU/ | Est. GPU | Max RAM| Memory | Neural | | | GPU Cores | FP16 Compute | Config | Bandwidth | Engine | +----------+-----------+--------------+--------+-----------+----------+ | M1 | 8C / 8C | ~5.2 TFLOPS | 16 GB | 68 GB/s | 11TOPS |

Comparison of apple M1-M5 Chips at max ram configs. M5 Max is not actually that good in raw float16 (same as M1 Ultra lol) but should go 2x+ faster in int8. Every single ai websites deep research results hallucinated wrong numbers while making this ( grok, opus, qwen g3.1pro), fixed it manually😠: