bearjaws

48 ms·

781.
by bearjaws 3y ago
1024x1024 resolution is a huge upgrade. I believe LLAVA is only 500x500 and it shows, its ability to even say what is inside an image is limited. Heres the chat version on HF: https://huggingface.co/deepseek-ai/deepseek
782.
by bearjaws 3y ago
Not really, the hardest part for most growing startups is scaling teams, not hosting.
783.
by bearjaws 3y ago
Seeing models like this work so well gives me hope that mobile first LLMs for things like better voice to text and typing prediction will not just 'work' in 2-3 years but actually not kill your battery too.
784.
by bearjaws 3y ago
Just need to start sinking ships, sadly. The rate of enforcement is far too slow, and you have far too many nations as bad actors, there will never be effective enforcement.
785.
by bearjaws 3y ago
A 100GiB bandwidth credit today is like Apple offering 8gb of RAM on the Mac Book Pro base model.
786.
by bearjaws 3y ago
I am most interested to see if the M4 chip will dramatically improve performance of local LLMs. I like M1 Pro so far with models up to 30-70b parameters, but the memory bandwidth is my current limit. With a large jump in unified memory and
787.
by bearjaws 3y ago
And of course the price minimum will be what we pay today, it will only get more expensive.
788.
by bearjaws 3y ago
Optum has billions of dollars in profit annually, none of that is going into the employees pockets or patient savings.
789.
by bearjaws 3y ago
Oh I don't disagree that the system is fundamentally broken, but we clearly need to attack from all angles.
790.
by bearjaws 3y ago
Go after all the healthcare giants. If people knew even half of how consolidated their treatment was, they would riot. For those who don't know about the significant conflict of interest present here: 1. United is a health insurer that
791.
by bearjaws 3y ago
I feel like Svelte has reached my ideal state in terms of framework. Building a intranet with it feels unfair to our main app. It literally loads instantly, the image content on our site is larger than the entire markup and JS delivered. Cr
792.
by bearjaws 3y ago
Surprised its that "cheap", almost certainly not profitable. I wonder how many organizations will use it, given you have to buy a license for each employee, that adds up quickly.
793.
by bearjaws 3y ago
Microsoft hands money to OpenAI, who then combines it with VC money to hand back to Microsoft. Same goes for AWS and GCP. They are going to take as much VC cash as possible and convert it to TPUs and electricity.
794.
by bearjaws 3y ago
I am not going to go into every nuanced issue, but to give context rounding errors alone costed us hundreds of thousands of dollars in customer support and eng time. I am talking Office Space style fractions of pennies type problems. Mostly
795.
by bearjaws 3y ago
Worked for a large travel website that rolled its own billing... Absolute nightmare when we blew up, had to hire 6 people to beef up the billing side of the website and it definitely cost us way more than those 6 people due to mistakes. I e
796.
by bearjaws 3y ago
Had this happen (but more like $100) analyzing the GitHub dataset on GCP. Honestly I was already concerned when it was taking more than 5 minutes to return a result. Once I saw how slow it was I found my error, not querying the sample datas
797.
by bearjaws 3y ago
EZ rage bait. Everyone of these platforms is scraping your inputs to train their next model. How is it not obvious they would also detect abuse?
798.
by bearjaws 3y ago
Nothing really wrong with FP8 IMO, it performs pretty damn well usually within 98% while significantly reducing memory usage.
799.
by bearjaws 3y ago
I believe that is where they are implying they do it without increasing memory utilization dramatically. If 1M context uses 32x the memory of 32k, its a non-starter. Even a smallish LLM like Mixtral uses 4-8gb of memory just for your prompt
800.
by bearjaws 3y ago
I wasn't worried until Google showed up with a 1M token context with 99.9% recall...
801.
by bearjaws 3y ago
The problem is, the 128k window performed terribly and showed that attention was mostly limited to the first and last 20%. Increasing it to 1M just means even more data is ignored.
802.
by bearjaws 3y ago
Cannot emphasize enough, even with the improvements in context handling I imagine 128k tokens costs as much as 16k tokens did previously. So 1M tokens is going to be astronomical.
803.
by bearjaws 3y ago
I'd argue it's about the same, maybe slightly stronger focus on build quality. Bought the studio buds plus and they don't hold a candle to Airpods. Gee I wonder why Beats doesn't have a direct competitors to Airpods?
804.
by bearjaws 3y ago
I give it 5 years max until enshittification takes over. Once they reach parity with CircleCi & GitLab, it will just be continual price increases. Bitbucket is basically a non-competitor and Gitlab is focused on reaching profitability,
805.
by bearjaws 3y ago
The last 5 Pixar films are all flops, and all signs point to leadership at Disney interfering with the creative process. > Also many bank acquisitions that prevent failure are obviously good. We wouldn't worry about bank failure if
806.
by bearjaws 3y ago
Clearly worked enough with Onn brand for them to keep consolidating. Just at the surface level this hurts: 1. Other brands selling TVs in Walmart. 2. Advertisement space will be cheaper for Walmart when they advertise on their own TVs, toei
807.
by bearjaws 3y ago
Shouldn't have ATT and VZW in the first place... Should be much smaller, in fact ATT was already broken up but then FTC policy changed. Bit more complicated due to spectrum licensing, but theres a way to make it work with more carrier
808.
by bearjaws 3y ago
I can assure you, every founder I've worked with would go public before kneecapping growth or value. It's not like when you sell your company you actually get all your equity as cash, you are certainly still working at your compa
809.
by bearjaws 3y ago
Assign domains of the app to developers, the people doing the code reviews require people to update documentation. Bakes into your existing workflows, direct line of accountability to both the code reviewer and the dev. The editing software
810.
by bearjaws 3y ago
The FTC just needs a new law based on anti-trust that says multi billion dollar companies can't buy other billion dollar companies. Period. Name the last time a multi-billion dollar merger worked for consumers.

More ›