Just as ChatGPT is now called "Chat" by the kids (go ahead and finish grinding your teeth, I'll wait...), if Antigravity takes off then it'll get an irksome nickname like Antigrav or just Grav.
I always wonder why not more companies are using more creative approaches like Google's Dutch auction to set their IPO price.
It seems direct listings gained some popularity but overall most companies seem to rely on the traditional underwriter model.
According to [0] -
> 22 companies went public on major exchanges using IPO auctions in the U.S. between 1999-2008, but there have been none since then, as of May 2025. Starting in 2018 when Spotify went public, there have been at least 20 companies that have gone public using a direct listing. With both IPO auctions and direct listings, underwriters do not have discretion to allocate shares to their preferred clients.
The Dutch auction wasn't all that smooth, with Google having to adjust prices at the last minute. Most results show that it didn't really produce all that much of a benefit over a traditional IPO, where (the issuing company) at least gets a guaranteed minimum amount of money to raise. Pretty much the only alternative is a direct listing for companies that want to go public, but don't need to raise money. (Well, there's also SPACs, but that's a different beast).
When banks, pension funds and other institutional investors are fighting to outbid each other to get a slice of your IPO why would you care about finding an accurate value and being "fair"? Whether the price goes up and down post IPO isn't really the company's concern. They already have the cash.
Google's process produced a price of IIRC $85/share and on first-day trading soared to >$100 so there was a lot of discussion at the time about how efficient that process was.
But really that's how all IPOs work, basically. You have one of more investment banks that underwrite the offering. They're basically guaranteeing to sell a certain number of shares to their clients at a given price. Those clients can be institutional investors, pension funds, high net worth individuals and so on. But there's a feedback loop here where clients might push back on a certain price.
IPOs love these sorts of investors because they tend to buy and not sell. If everyone sells the IPO will flop. Retail investors are far less "loyal".
The IPOing company also has levers where they can manipulate the price, most notably on the supply side ie by limiting or expanding the size of the float. SpaceX's float (as a percentage of the company) was actually really small.
What's unique about the SpaceX IPO was that it would immediately become one of the world's most valuable listed companies so there'd be a lot of induced demand from index funds. The underhanded (IMHO) aspect to all this was that the rules were deliberately changed so passive investors would be exposed almost immediately rather than first allowing some form of price discovery by the market. NASDAQ capitualted. S&P did not.
I guess the real manipulation here is the fiction that SpaceX is an AI company, which ultimately goes back to a series of bailouts for terrible decisions going back to the Twitter purchase. SpaceX's AI pitch was orbital data centers, which make no sense, and using their ill-gotten NVidia chip allocations to rent them to Google.
> In regard to AI usage, my fuzzing workflow was automated by AI with a strict harness. I used GPT-5.5-3-Codex-Spark for ALL the fuzzing, as barely any "thought" is necessary when provided with an efficient harness. Contrary to the growing narrative that I'm just some random child burning tokens, I DO actually have a degree in the subject and have published multiple papers on fuzzing methodology. I spent years researching and developing new tools and ideas for how to fuzz. You do NOT need a SOTA model to help you identify these issues, I promise!
If this project is an ad for their product (Upstash, promising "Highly Available, Infinitely Scalable"), then the last thing they'd want is a hug of death :/
Our startup (~20 people) got slashdotted in 1998 or so. I was the only one randomly awake at the time. Remember watching all the logs from our web server in realtime, ready to immediately kill anything or anyone threatening the overall availability.
512 kbps uplink, I think. Even accidental DoS was trivial. We had a self-hosted little data center at our office with the only available stupidly expensive commercial connection.
Felt some dread having to restart the main (async, single-process) web server a few times to keep things going due to bugs in our code. So many* people on dial-up patiently waiting for the page to load.
Its funny that these days the bottleneck is usually the data layer. Servers are so powerful now that even your average $5 server can handle HN levels of load if configured correctly.
/api/hn -> 502 {"error":"Search entry should have an initialized schema, command was: [\"SEARCH.AGGREGATE\",\"hn\",\"{\\\"$or\\\":[{\\\"title\\\":{\\\"$eq\\\":\\\"anthropic\\\",\\\"$boost\\\":5}},{\\\"text\\\":{\\\"$eq\\\":\\\"anthropic\\\"}}]}\",\"{\\\"by_month\\\":{\\\"$dateHistogram\\\":{\\\"field\\\":\\\"time\\\",\\\"fixedInterval\\\":\\\"30d\\\"}},\\\"top_authors\\\":{\\\"$terms\\\":{\\\"field\\\":\\\"by\\\",\\\"size\\\":6}},\\\"by_type\\\":{\\\"$terms\\\":{\\\"field\\\":\\\"type\\\",\\\"size\\\":4}}}\"]"}
I'm also getting /api/hn -> 504 An error occurred with your deployment FUNCTION_INVOCATION_TIMEOUT cle1::48fnt-1782412720840-4855b2b75b5a after a few lookups
I've been working on a pure Clojure implementation of WebRTC Data Channels (SCTP over DTLS over UDP). The library provides a minimal, dependency-free (except for Clojure itself) way to establish peer-to-peer data channels on the JVM.
I've always wanted this and have used it to experiment with Gemini's cloud agent Google Jules.
Thanks! I've noticed a big jump when they switched to Gemini 3.1 Pro and it really became useful. I like that I can use it from my phone too. It took a bit of trial and error but I came up with a good ralph loop between GitHub Actions and Google Jules using the Jules API. So basically I have Jules extend its TODO.md with the next set of tasks and open a PR then run a GitHub Action with a few checks, auto-merge, and then call back into Jules to kick off the next cycle if there are still open tasks. It then mostly just runs and occasionally gets hang up on some questions that I then answer on my phone mostly just telling it to make a judgement call and keep the build green. You can check out the prompt, action, and past PRs for examples ex. Jules prompt is here: https://github.com/alpeware/datachannel-clj/blob/main/prompt...
Very cool, thanks! I have an under-utilized Google AI Max plan and this has inspired me to investigate Jules further (and Antigravity for that matter). It's been a while since I've checked in with Google's AI suite.
I have an upcoming project in Flutter, and maybe it's wishful thinking, but my intuition is that perhaps Google has top-tier LLM performance within their own ecosystem relative to peers.
the model doesn't know itself, but all these larger models are generating a significant amount of synthetic data from the prior models, and the prior models are all context bloated renditions; you fill the KV cache with whatever alignment you want, and then generate synthetic data.
That training on existing models is what brings out various other things about other models; then there's models that are just like snowballs, where you build one iteration, then you give it it's identity, then you train on that with the same synthetic generaiton.
So a model could generation include at some point it's own name.
I don’t think what you’re saying makes a lot of sense. You don’t “fill the KV cache with whatever alignment you want.” That doesn’t exist. The KV cache is an inference optimization, and is populated by running tokens through the model.
Synthetic data is generated by other models, and yes this is often where identity propagates.
I think with the snowballing you mean things like iterative self distillation? That’s definitely not done unsupervised, because of the risk of model collapse, and typically heavily curated and/or mixed with real data.
From their blog post announcing the platform end of August 2026 [0]
[0] https://www.appliedcompute.com/platform/introducing-ac2
reply