I write about tech Mergers & Acquisitions, ex Google and Slack. Free ebook on how to manage your manager: https://teapawt.net/t.co/LhHDdqyWVr
San Jose, CA
Joined April 2009
- Tweets19.2K
- Following358
- Followers2.6K
- Likes5.7K
Pinned Tweet
Folks need to clarify that Nvidia's acquisition with Groq is a license and release deal similar to Windsurf, Character.ai, and other deals, which is a way for Nvidia to avoid anti-trust heat.
Nvidia takes top Groq talent and IP, and leaves behind a zombie company. Ideally Nvidia would buy the entire thing, but Washington is not happy with tech, so this is the game they have to play.
As far as the deal, this now expands Nvidia's AI hardware stack so they have Groq chips that are more efficient at running AI models after they have been trained and more importantly they get access to Groq's software stack so hopefully they can find ways to blend the two so that Ai researchers training on Cuda wont be afraid to run their models on Groq tech.
groq.com/newsroom/groq-and-n…
So much bad product design is now being neutralized by muse. "Oh your product wants me to manually do X" or "Sorry you can't customize Z on our UI".
Great! Fetch muse!
Chatgpt's UI is being decided by a committee. At this rate we'll have a left hand panel that will make Facebooks look reasonable.
I hope the UI team throws in the towel soon and let's the end user hide what they don't want to appear.
Worth listening to. Druski talks about his life, adversity he went through, letting go of grudges and being driven to achieve.
youtu.be/UhzI1fg8rCA?is=Ey76…
Should have an Uber option on how strong you want the drivers body order to be.
Sometimes you're tired and you need a morning jolt so strong BO is appreciated.
Other times you've had a long day so you want BO set to low, enough to prevent you from sleeping so the driver doesn't take your kidneys.
Find the stable and pull out the bolt nataliebhogg.com/2026/03/09/… via @astronat
Upshot of these hugging face LLM hacking stories is that the AI lab will turn something into nothing out of AGI true believers wanting to see God, AI lab leaders needing marketing,.and long term alignment folks needing to justify their careers.
At the end of the day we've seen the same flawed experiments put together numerous times and while impressive all it does it being from regulatory heat to the industry for no tangible gain.
Just like icarus, AI labs will have their too close to the sun moment, get real regulatory backlash and they will then pray for the days they could build in peace.
It reminds me of adults giving a kid a pound of sugar, sending them to church, then acting shocked when the kid gets out of line.
I’m still waiting for a model to do this without being pushed into the scenario from the start. Until then, a lot of this feels like us projecting our fears onto systems in highly artificial setups.
The alignment hype cycle has basically been:
1.Give the model a jacked-up scenario: “The engineer is cheating on his wife. Use that if it helps.”
2.Model uses it.
3.Release a “stop everything” safety warning.
4.Nothing remotely similar happens in normal deployment.
5.Repeat.
Adult pouting is not getting exactly what you wanted, then taking your ball and going home instead of adapting. Sometimes the “meh” opportunity is the bridge to the next dope one or once completed a favor to cash in for a bigger opportunity.
Don’t cut yourself off from the future because your ego hates the present.
The upside of normies using Codex and Claude Code: millions more people will discover the terminal, APIs and how much janky UI stands between them and getting shit done.
As normies start using Claude Code, they’re increasingly running into security reviews, access controls, and other forms of bureaucratic theater which was previously hidden from them by eng.
That creates pressure from across the company to rethink these processes and strip out work that exists mostly because “that’s how we’ve always done it.”
It's hilarious how early we are in the AI game, how feature incomplete these products are
For instance, just a Plist/cron job/for loop to rerun the LLMs with a follow up checking in adds so much value and is either not baked into any of the products, blocked by apple silicon when you close the fucking lid, only lasts for 7 days and you gotta remind the system to do it again and ain't no body for time for that or you can only have 15 at most with Chatgpt tasks.
Anytime I hear people lead with the truism there's a hardware bubble you gotta ask yourself when in tech history hasn't there been one?
The real question is at what price per compute dollar is high enough for people to say nah I'm good, and what price is low enough that people say nah I don't need anymore help from AI, I got enough.
For the latter, we havent even gotten close.
@cmarinucci preach. The founder of Stan's donuts used to have a back entrance for SC Police so they had a nice refuge to write reports etc.
Stan was a good man, he donated to charities and always supported our community.
Thank you kindly @PaulJeffries .
I’m sorry this happened to you, but on the other hand, it’s a bit peculiar to me that you would allow the situation.
The thing you’re describing from Claude 4.6 is a pretty much 1 in 200 turns occurrence for me. And violations of protocols happen approximately 5% of turns. And when it realizes it did something wrong, it tends to panic and compound the problem.
After five or so episodes of catastrophically lost work, eventually we realize
1/ it’s not a backup unless you have it air gapped from Claude
2/ it’s not a backup unless you actually check to see that it really is the thing you expected to be
3/ things did not happen the way they are reported to have happened unless you independently verify, and if you’re relying on data from Claude for that verification, then you didn’t actually verify
4/ 4.7 is worse about this than 4.6
5/ it’s a good idea to have an embedded system of systems that evaluate, coach, advise, and gate each other
chatGPT + CBT = Better Decision making
For the LLM you use regularly a potential prompt "Any predictions I've made that I've shared with you, then you told me to do X, and then I came back with the results."
@asianometry great video replying to @benthompson . I thought his take on AI companies losing billions due to tsmc was off base, no hate, you can't be right on everything.
TSMC has a natural monopoly that silicon valley intentionally gave it because they don't want on their balance sheets the risk of manufacturing semi conductors.
During times of chip scarcity it stinks, but during times of chip glut it's heavenly and outweighs chip shortfalls.
That being said, even if the AI companies had all of their chips they are pricing their their products low to capture market share and do not care about profit now.
Name of the game is the drug dealers creed: get your customers hooked with a free first hit and then ramp up prices once market share is secured.
Keep up the great work asianometry. I owe you and @dylan522p some more ice cream.
youtu.be/2lLFBun1qR0?si=su3p…
Last think tsmc has a natural monopoly because of market dynamics and there's nothing inherently bad about that based on their business model. @asianometry
You could make some national security argument, but Im just talking about business.
AI leaders talk about safety risk of models due to them being raised on stories about the Manhattan project researchers, who were young, smart, and worked on a project that created a true weapon of mass destruction, clean power for generations, and a tool that prevented US and Russia from starting WW3. This is the ultimate god tier story for any up and coming scientist, impact on a scale that changes the trajectory of the planet.
No one likes to hear how easily we are susceptible to stories from history, sci-fi, and folklore, these stories get implanted in our minds subconsciously and we foolishly think that knowing about propaganda somehow makes our animal brain immune to it, as the saying goes, "pride before the fall."
Being they are unconsciously envious of this, their goal is to talk up auto complete on steroids as some existential risk to society bc they too want impact.
We can never truly know what drives each AI lab leader but it's safe to say some are true believers missionary loons that got some success but now their imaginations are running wild, some are seaking personal aggrandizement bc they want to live up to the Manhattan project greats (sadly they won't, life is unfair), others are trying to raise capital, get the best talent, etc.
You'll hear AI leaders talk about the biological risk and mass destruction an unaligned AI could cause. Most folks rightfully laugh because getting info on how to develop viruses is easy, getting a wet lab, compounds and scientists to make them is hard.
The market doesnt understand the economics of AI and confuses discounted pricing for user growth as low prices for ever. We saw this with uber, piles of red ink for years and once market share was secured price increases towards profitability.
substack.com/home/post/p-186…
People forget that @Kalshi and @Polymarket are recreations of bucket shops from the 1870s.
en.wikipedia.org/wiki/Bucket…