R1 was a frontier model when released this January. GPT-OSS-120b is only a couple of weeks old and I believe is on the cost/capability pareto frontier on some classes of benchmarks. (Though to be fair, GPT-OSS was trained almost entirely on synthetic data and heavily benchmaxxed for math and reasoning, so it seems that not a lot of people are finding it to be that good at real world tasks.)
It seems that at this point nobody except xAI is trying to brute-force the problem purely by adding more scale to the pretraining or by just adding more scale to the reinforcement learning. Iâm sure theyâre all spending all the compute they can spare on training, and will definitely be spending a lot more on it in 2025 than 2024, but itâs being spent more smartly than âspend 10x more to get 10% improvementâ.
That seems incredibly implausible, given they were as well connected with SV VCs as is humanly possible, and had $1B in initial funding. They would have no need to find retail investors through some tiny Swiss private market. Crunchbase also knows nothing about these multiple seed rounds.
This is the first Iâd heard of Moonshot, but from a little reading it seems that these were not deals for actual shares but at best some kinds of derivatives with non-credible counterparties. But more likely a ponzi scheme.
Moonshot had a tracker certificate (or at least on paper they had) for a lot of PE oppties (SpaceX, anthropic, openai, etc) on obscenely high fees. Could have been a win-win.
Maybe it was a scam, we will know in some months when the bankruptcy agency in Zug is done.
I think âare we in an AI bubble or AI bonanza â deserve a separate thread
i feel it would be both
AI will change our lives but we are also over estimating its impact and utility. I am not working in IT companies so I donât have inside view. But I work in industry and I can say for now the use cases are not at scale except summarisation , content creation and task management.
And to be honest looking at all the AI slop, I hope the content creation via humans still have a place or else we will need to read through 1000 pages of garbage to get to 1 page of real information.
I was talking to someone from Google and they told me that when google says X% of the code is written by AI, it is a bit of exaggeration because it includes certain things like auto complete etc which was already in use for some time.
Most of what weâre building out at this point is the inference [âŠ] Weâre profitable on inference. If we didnât pay for training, weâd be a very profitable company
A little bit like a car manufacturer saying
âWe are profitable on sales and service of our cars.
If we didnât pay for developing and manufacturing them, weâd be a very profitable company.â
A more apt comparison would be to say âWeâre profitable manufacturing and selling our current cars. If we didnât develop new models weâd be profitable.â
Looks like the US leaderâs quest for âutmost loyaltyâ doesnât stop at government employees. Outreach to C-Suites: your well-behavior (as defined by us) will matter:
The West Wing has created a scorecard that rates 553 companies and trade associations on how hard they worked to support and promote President Trumpâs âOne Big Beautiful Bill,â a senior White House official tells Axios.
Why it matters: Trump works transactionally, and companies have rushed to pay demonstrative homage. Now, senior aides will have data to consult when considering corporate requests.
Iâm surprised about how many respondents have their company pay for a subscription (more than half who use it if Iâm interpreting this correctly).
Iâm also surprised about the 16% not using AI directly. Especially given the sample bias I thought weâd be close to 100% (direct) usage of AI.
Is this result a sign that the forum is less smart than I thought or smarter than I thought ⊠[]
Itâs even harder to judge as I believe the technology is a big step forward, but as as a short-term 4-5 years time horizon investment I remain a skeptic.
Also, folks, someone tell the guy who set up the poll that itâs spelt âneitherâ not ânetherâ. Sheesh!
Maybe the guy should use AI to set up polls with better spelling âŠ
30% of the forum population are Google employees with presumably access to Google AI (or, for a few, software engineers/data scientists in a start-up with their own model).
16% are finance/pharma people, lawyers, doctors or civil engineers, who are notoriously slow to adopt new tech.
âLess smart than you thoughtâ or âsmarter than you thoughtâ is only a reflexion on the limitations of your own thought process. You should ask AI. Ideally several of them and then ask some of them to compile their common reasoning and then a few to assess the validity of the compilation. Or you can just trust the one with the most sexy in-built voice.
I replied that I donât use any AIs, but then realised Meta forces its AI on WhatsApp, Google forces its one in Google Docs and Microsoft forces its Copilot on my work email system.
I donât actually actively use any of them (if I wanted inaccurate information presented confidently I would just ask certain colleagues), but they all probably count me among their âactive monthly usersâ as I might accidentally open them occasionally.
Mit dem Lesen und der Teilnahme an diesem Forum bestÀtigst du, dass du die Forum-Richtlinien gelesen hast und damit einverstanden bist sowie den Haftungsausschluss auf http://www.mustachianpost.com/de/ akzeptierst.
]