Hansenq

Historically, when a company gets acquired, the terms of the acquisition vary wildly. Many acquisitions over the past few years have led to a payout of all equity holders, pennies for employees, and layoffs for much of the existing team. This outcome is no worse than an existing acquisition--people just want details because of the new structure. My point is that this is financially no worse than an existing acquisition. It just _feels_ worse because of how it's structured.

And to your last paragraph, read reference [1]. The distribution of that 2.5B is in accordance to the existing cap table; it will make sense once you read that tweet. You must allocate money according to the cap table, and so that allocation is already determined in a company's previous funding rounds.

This is the logical, satisfying, and probably best conclusion to an un-ideal and optically terrible situation all parties were placed in.

First, OpenAI wanted to acquire Windsurf. Terrific move! Win-win for OpenAI (who needs more AI product) and Windsurf (for the deal price). But this fell apart because Windsurf didn't want the IP to go to Microsoft (which imo should not have been not a big deal, especially if you knew what would have happened next). Big loss for all parties for this to have fallen apart.

My biggest question still is why not continue on as an independent company? Perhaps losing access to Claude doomed signups; perhaps employees/investors had a taste of an exit and still wanted it; perhaps due to fiduciary duty to maximize returns; perhaps their growth stalled due to the announcement? In any case, the founders got a similarly equivalent deal from Google, and were arguably wise to pursue it.

But Google's Corp Dev team here is the most maddening. Why not fully acquire the entire company, instead of doing the same "acquihire and license" deal that was done to Character AI, Adept, Scale, etc.? Risk of FTC antitrust review is a thing, but Google's not even competitive in the coding market, so I doubt there is a review (though I do hear that all acquisitions by large tech companies these days are reviewed by default). If there's anyone to blame in this situation, it's the FTC and Google for pursuing this strategy, instead of a full acquisition. Win-win for Google (for the team) and Windsurf (for getting a similar acquisition price, but liquid!).

Imo, the founders did a good job ensuring that close to the $3B acquisition price was reflected in the $2.5B Google deal--all existing investors and vested employee/equity holders are paid out; the company also retained $100M which was suspiciously similar to the amount needed to pay out all unvested employee/equity holders [1]. So theoretically the remaining company could pay accelerate vesting, then pay out the cash to their remaining employees, and then shut down, to give everyone the same exit as an acquisition, or better. This might have been the best scenario, because the brand damage to Windsurf as an IDE that happened over the weekend was pretty close to unrecoverable for them as an independent company.

But instead, the company leadership decided to field acquisition offers for the remaining company and IP, and got one from Cognition. (I'm actually surprised this acquisition isn't under FTC review; it's more plainly an agentic coding company acquiring a competitor agentic coding company). In taking the offer, it reinforces that the Windsurf IDE will continue to exist, that they have a R&D team backing the IDE again, and can marry Windsurf's enterprise sales chops with Cognition's product [3]. Win-win for both Cognition and Windsurf.

So overall, win-win-win all around, except for OpenAI, Varun's public reputation (imo, undeserved), and startups hiring employees (who might think they might not get a proper exit) [2].

[1] https://x.com/haridigresses/status/1944406541064433848

[2] https://stratechery.com/2025/google-and-windsurf-stinky-deal...

[3] https://x.com/russelljkaplan/status/1944845868273709520

I'm surprised most commenters haven't mentioned that the presence of Tiktok as the biggest reason why Facebook was pushed into this direction.

Ben Thompson of Stratechery did a great deep dive into Facebook's Three Eras here (https://stratechery.com/2025/meta-v-ftc-the-three-facebook-e...). Essentially, Meta could afford to prioritize positive well-being when it had a monopoly on social media, but as soon as Tiktok came onto the scene and Meta started bleeding users to it, they had to respond. Now, everyone (Instagram, Youtube Shorts, Twitter, LinkedIn) is copying the model of vertical auto-scrolling short-form videos, because it's a battle for attention.

What _was_ Facebook supposed to do when it saw all of its users leave Instagram/Facebook for Tiktok? Not do anything? Though it's terrible that everything is now a short form addicting video platform, I understand the logic behind why the company did what they did (and why everyone is building this). People say they want real connection, but really, they just want to be entertained.

React is used to render the CLI through ink: https://github.com/vadimdemedes/ink

Yoooo congrats Micah!! Joining Brandon and Brayden as W23 acquisitions by Cloudflare!

Maintaining these escalators must cost a fortune, not only because it's an old model that probably doesn't exist anywhere else in the world, but also because the US elevator market is completely distinct from the rest of the international market. This leads to higher prices since parts can't be shared.

https://morehousing.substack.com/p/elevators

This is one of the clearest example of overzealous regulation leading to the West losing one of its strongest household robot brands to Chinese upstarts.

Robot was not in a particularly financial good shape beforehand, but at least had models that could compete against the Chinese competitors. Because it was running out of cash, it tried to get acquired by Amazon for $1.4B. This was, by no means, an ideal deal, but it would have put the company on stable financial footing, and let it continue as a brand (similar to Alexa or Ring).

However, European and FTC regulators prevented the deal, saying that this was an invasion of home privacy by the big-bad Amazon. Fast forwards three years and today, all of the robot vacuums you buy today are Chinese-made and iRobot is about to declare bankruptcy. We traded off home privacy from Amazon to an upstart Chinese company and lost a leading $1.4B company. Great. I don't know about you, but I trust Amazon more with my data than a Chinese company.

https://www.reuters.com/markets/deals/amazon-irobot-abandon-... https://www.theverge.com/2022/9/29/23379036/amazon-irobot-ro...

I wonder how much of these conclusions are Claude-specific (given that Anthropic only used Claude as a test subject) or if they extrapolate to other transformer-based models as well. Would be great to see the research tested on Llama and the Deepseek models, if possible!

Thanks for the feedback! The term "PRD" is pretty overloaded in that it's often used interchangeably by engineers to refer to technical design docs/requests for comments/spikes/etc (people really need to align on a single term for these), so I used it here to capture everything. Agreed that we didn't build the product as it currently stands for PMs--working on integrating other sources like GDrive, Notion, Coda, etc in the future though!

On your second point, are you saying that the quoted text sound like AI slop? Interesting feedback...I hand wrote that without AI so perhaps our prose style is different. Will give it more thought.

GPT-4.5 is 15-30x more expensive than GPT-4o. Likely that much larger in terms of parameter count too. It’s massive!!

With more parameters comes more latent space to build a world model. No wonder its internal world model is so much better than previous SOTA

Many of those articles came out before TSMC received CHIPS Act grants. As soon as the CHIPS Act money was committed to TSMC, the factory was suddenly ahead of schedule. Noah Smith called it out here:

> Three months after TSMC announced further delays at its $40 billion Arizona fabs, the chip manufacturer has now said the plant is expected to be operating at full capacity by the end of [2024].

> The announcement comes several weeks after it was first reported that TSMC is set to be awarded more than $5 billion in federal grants under the US CHIPS and Science Act…

https://www.noahpinion.blog/p/at-least-five-interesting-thin...

Reading through the Chain of Thought for the provided Cipher example (go to the example, click "Show Chain of Thought") is kind of crazy...it literally spells out every thinking step that someone would go through mentally in their head to figure out the cipher (even useless ones like "Hmm"!). It really seems like slowing down and writing down the logic it's using and reasoning over that makes it better at logic, similar to how you're taught to do so in school.

Props on them for fixing this. Spent all day tracing down why my deploys were failing; turns out it was way way way upstream of my code, into the latest node docker container facepalm

an available .com is basically the only reason you should use https://paulgraham.com/name.html

I used to go down rabbit holes on Wikipedia all the time--could spend hours doing this.

This to me seems like an Infinite Wikipedia! Really cool use-case!

If you read the WaPo article that's the topic of this thread, you'll see that the actual voice artist is quoted in the article.

I think OpenAI is targeting clear enterprise use-cases for what they're building into ChatGPT. Data Analysis is a clear enterprise use-case. So, if a feature helps them sell ChatGPT Enterprise, I think they'll build it, since that's a large revenue driver for that.

Consumer-focused wrapper startups like jenni.ai, research helpers, or math tutors I doubt they'll focus on, since most of the revenue is in enterprise.

The way ChatGPT Data Analysis works is that ChatGPT generates python code that does the transformation step. Because python code is static and deterministic, you can always re-run the code to get the exact output over and over again. If you want it rolled back, just re-run the code, but run fewer lines of it.

In fact, this is probably more accurate than custom UIs that log actions you take, since to build that UI, engineers need to make sure each action is logged (and it's easy to forget to log one!).

If you're looking for a more powerful version of ChatGPT Data Analysis, I recommend https://julius.ai/.

If you're looking for a version of ChatGPT Data Analysis, but want to offer it to your customers, take a look at us! https://www.lightski.com/ (disclaimer: I'm the co-founder.) We're building this feature of ChatGPT, but in a way that you can embed inside your app, so that your end-users can get an AI Data Scientist without needing to export data, then import it into ChatGPT or Julius.

These are called double-loaded corridors, which are basically buildings required by North American fire codes to have two accessible interior stairwells for each unit. Unfortunately due to architectural constraints, this means that all new buildings built in the past few decades have interior "hotel-like" corridors and exterior flat, boxy frames.

North America is the only country that requires this type of fire construction--most other countries allow one interior staircase and an exterior staircase out of a window (like a ladder). Without this requirement, you're able to build buildings that are shaped differently, have windows in multiple directions, better support 3/4/5 bedroom condos, and are cheaper.

There's a movement to harmonize our fire codes with international standards; you can read more here: https://www.centerforbuilding.org/blog/we-we-cant-build-fami...

If people are looking for additional in-depth reading on how this happened, The Air Current did a great write-up on this systemic mistake using internal Boeing sources a month ago that the NTSB report fully supports: https://theaircurrent.com/aviation-safety/127-days-the-anato...

ooo thanks. Too late to edit now but appreciate it! Not a pilot currently; just very interested and will probably get one in my lifetime.

There's a follow up video to the one linked in the story that provides a lot more context missing in this piece. https://www.youtube.com/watch?v=4zHxdn8oz20

Basically, SFO normally does VFR parallel approaches at night. Approach sequences these approaches miles beforehand, so there can be a chain of 10-20 aircraft all sequenced to land before responsibility is even transferred to SFO's tower. The incident happened during a particularly busy landing time at SFO, so there was indeed a massive chain of aircraft coming in to land.

Lufthansa was the only aircraft asking for ILS. Because ILS needs greater separation, that would require breaking the chain of approaches, sequencing a single ILS approach, then resuming. The chain of landings already sequenced takes priority, so Lufthansa would have to wait 30+ minutes for a gap to appear. By the time that gap appeared, Lufthansa had just decided to divert to Oakland. If Lufthansa had arrived a bit earlier or a bit later, they would have been sequenced just fine.

ATC could have been a bit more accommodating in rerouting their divert to SFO as soon as the a gap appeared, but Lufthansa was also the only airline requesting ILS, and they're already dealing with sequencing 20+ aircraft during a busy time. It's not clear who's in the wrong here; just an unintended consequence from many well-intentioned decisions.

I tried out Monaspace but felt that the fonts were a bit thin for my QHD monitor I use as my primary display. It might be better on a retina/4k monitor though.

Perhaps fonts are something you get used to after some time using it, but I ended up switching back to my favorite font, Input Mono (which, as a coding font, isn't actually monospace, so it brings a bunch of cool features and doesn't need to do texture healing).

https://input.djr.com/info/

Congrats Steve and Andre! Product has come a long way since I got a demo 9 months ago, and definitely see the applications and applicability of the product now :)

Looking forwards to what you guys are able to continue building in the future--and will likely be using Valtown for scripts of my own!

I feel like the biggest takeaway here is that a classifier trained on samples could only predict whether or not ChatGPT would refuse their response 76% of the time, which to me seems very low (given that they used BERT, regression, and a random forest as their classifier).

Probably means there's a lot we still can't predict about how LLMs work internally, even if we try to apply classification to it.

Congrats on the launch! We built something similar a few months ago but couldn't get the quality of the SQL generation to be that good without providing a ton of few-shot examples, especially with databases that contained more bespoke data. If you have enough few-shot examples the quality might be quite good!

Congrats! We've been happy users of Helicone for the past few months--it literally helped us solve a bug with OpenAI's API where we didn't know why requests were failing and we failed to log some of their responses. Helicone helped us debug that it was a token limit issue really quickly, especially since the logging around hasn't been great.

Love how easy it was to integrate too--just one line to swap out the OpenAI API with theirs.

Is Mac's Preview app susceptible to this bug too? Right now it seems only confirmed for Pixel phones and Windows Snipping Tool, but I use Preview and I haven't seen anyone confirm or deny it for that one.

Love that you can open up a PR to immediately update documentation directly from Slack! Underrated feature; I hated updating documentation so that's an incredible change.