yalok

pretty wild that, per public sources, Reflection AI raised over $4 billion and hit a $25 billion valuation while operating in total stealth, without ever releasing a single public product until now. Beam (501B) seems be their first-ever model drop. Or am I missing something?

Amazing. Kind of reminds of the modern day discourse on potential dangers of AGI and humans misuse of it.

diesel fuel itself needs to be heated at -20 C and below. I used to have a diesel car which would start itself up periodically when its fuel temp was getting close to -20 C, just to warm up the tank & pipes.

if you get into standstill traffic in the winter, it's actually life-threatening not to have alternative power source that can warm up the passengers (and the battery). One interstate road I'm driving sometimes, has many abandoned Teslas sitting on the side of the road after a sleepless night of staying in traffic during some snow blizzard...

> macOS has long provided a simple means for apps to handle dictation and transcription in processes that stay securely on the device

Not sure these guys realize that the quality and latency of those Apple services in MacOS is way lower than SOTA and not too many people use them because of that…

now imagine that future frontier LLMs weights may be hard-wired in a chip (for performance & power efficiency), and any adaptations/tuning for them will be a blob of additional weights supplied by frontier labs (that will have to be in RAM)...

you also need to factor in other things, like sentimental value (may be very high), rarity of some things, and catastrophic events "insurance" (e.g. Covid or similar events where obtaining some things is not guaranteed).

sounds like a perfect fit for ASIC-optimized models (where matrix ops could be supported directly in BITCOS format, potentially) & achieving record power efficiency for on-device inference.

And it looks like per [0], a model needs only ~30% more weights to be at comparable quality, if quantization-aware training is done...

0. https://arxiv.org/pdf/2402.17764 - The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

so this post was at the top of the 1st page of HN just a few min ago, and now it's down on Page 2 (#36 atm). Is this due to the title change, or something else?

this is unfortunately very common - I kept running very small ads regularly over many years (~10 years by now) and the % of bots has been steadily increasing, and Google doesn't really have any system to report those reliably to them. In fact, it's contrary to their incentives to investigate & fix these problems - unless they feel some pressure from competition...

From time to time, Google ads sales reps call me and are trying to convince me to increase the ads spend. I complain about bots, and none of them had any suggestion on what I can do to appeal it etc. I mean I can appeal on some specific instance, once - but there's no process to communicate back to Google the cases where I'm 99% certain about bots on regular basis.

And these bots get increasingly more sophisticated. They used to just click on ads. Then they started installing the app. Then running the app once and do nothing in it. Then they started tapping on the app screen & actually try to go through some initial app steps. When they do it in a spiky way, it's easy to detect them. But when they do it through some distributed device farms, below noise, it's very hard.

I was just answering a question above about the email access concerns. What you mention is another concern, but personally, if they are going to be showing ads to me, I'd actually prefer them to targeted for versus some totally random stuff.

As for training, I'd be highly surprised if they use the data they get from APIs for training. But the conversation itself & thinking trajectories probably will, unless you opt out, like you pointed out.

yes, it was not easy to mentally give that access, but then I heard from friends how much work was put in into securing it and separating LLM/agents from the API tokens access (also saw this detailed blog post from the lead developer - https://research.meta.ai/blog/security-and-safety-for-ai-age...), plus I had the pressure to get the problem resolved - so I decided to risk and try it out.

Eventually, though, Muse Confidential VM is an even more locked-in option coming, per that blog post...

I'm not sure - it may as well be that regular consumers just don't realize they need this and will greatly benefit from it.

My personal anecdote - we had a complex family travel in the summer, with 6 ppl and 6 separate flights booked (some multi-leg). SAS messed up on one booking - they decided to cancel the flight we had from Oslo, and rebooked us to an earlier date (2 days shift). That wouldn't work for the rest of our schedule, so I had to rebook, overpay extra money for now longer & more expensive flights - but to add the insult to the injury, when I did this through their website, they didn't transfer the food that I paid for the earlier reservation. Calling them in roaming and trying to get it fixed with a human support agent didn't help - the agent said he just can't fix it on his end, and suggested that we just file for reimbursement on their site.

I never had time to file it since July.

But today I had a reason - to test Muse - and asked it to handle it for me. To my surprise, it dug through all of the SAS emails and actually found the emails confirming that the food fees were reimbursed exactly the same date when SAS moved our original flight. Apparently I didn't notice those notifications since they were spamming with all kinds of notifications that day. But, if not some tool like Muse, I wouldn't even find time to handle this and either file a claim or figure out my miss on their earlier reimbursement.

first of all, thanks for building this - that's amazing!

Quick question - does it really need external SSDs, or if the local SSD fits the whole model - how fast the model would be? e.g. on your machine, M5 Max 128GB, with 4TB SSD? maybe it'd be good to add "0 external SSD" column on your graphs?

while these 1200 agents were fooling around to cheat on a benchmark and achieved impressive results despite of the limitations (sandbox, no internet, no intercom at first), one can imagine how much more efficient a similar army of agents may be in the hands of a malicious actor launching them without any of these limitations and with explicit encouragement to achieve some malicious goal at any cost... scary times.

and more specifically - what harness is known to be the best fit for Qwen local models, and are there any evals/benchmarks for harness+model pairs?

One can get kids doses that are 0.5 or 0.25mg. Eg see this 0.25 product on Amazon - https://a.co/d/0j74eWFP

Could star link add some cameras on the back of those satellites and make the detection actually much better?

is ultrasonic scanning completely harmless for developing baby? when my wife was pregnant, I remember they wouldn't recommend too frequent ultrasonic scans...

Impressive speed up at the cost of quality - a bit lower quality than 12B model, but multiple times faster…

if there're some specific tests/evals to satisfy that an agent can test by itself, it can easily iterate for hours. And this time also includes running those tests/evals, which may not be small.

0.8GB is for text only. It's more like ~1.1GB if you include video/audio encoder

Could just be more tests? :) Which is good for code quality in general and reduces support burden, but doesn’t lead directly to more features

imo, it really depends on what you enjoy doing. Regardless of AI, choose software development if you like to build complex systems that no-one has built before, and have enough patience to dig deep / debug things to make them work exactly as you expect.

For some of us here, it's just what we love to do, no matter what tooling is available. When I first started building my own software long time ago, it was a very slow Basic and fast raw machine codes (in octal system, PDP-11 like CPU). I enjoyed it not because of tooling, but despite of it.

Over the years, the tooling was getting better in general, which allowed us to build increasingly more complex systems.

With AI, we will still be creating & debugging. It's just that before AI, I had to spend 90% of my work on mechanical not-so-fun things to get things to work, and only 10% on fun algorithmic-intensive parts. But with AI tools, this ratio seems to change, and all kind of boilerplate code & algorithms can be written much faster by AI, hopefully leaving more time for us to work on creative part of the work.

just saw a relevant HN post about this very matter in Berkley - https://news.ycombinator.com/item?id=48392004

I heard they do CS exams on air-gapped machines at UC Berkley. Use of AI to do CS homework is strongly discouraged, and if someone cheated, it shows up at the exam...

but ASML is in Europe - so they hold at least some critical part of the stack.

> First git itself is distributed and built for scale.

there're different dimensions for "scale" - like handling large monorepos, orders of magnitude more commits, tighter requirements for latencies (for agentic use, e.g. for agentic history navigation)...

were you in that room where Adam was making that call? No? didn't think so...

Just give people some benefit of doubt. There're much simpler ways to explain certain things that suspecting some universal evil in every move...

There're tons of ways to fine-tune WebRTC that it wouldn't corrupt audio in poor network - it has all of the controls to smoothly trade-off latency vs quality. Not just NACKs - FEC, disable PLC/Acceleration/Deceleration, larger JB (tons of parameters) etc.

Most of the glitches I heard with OpenAI's Voice were not WebRTC related - but rather, to my ear, they sounded more like realtime issues with their inference - which is a very different component to optimize.