> Although that was only implied without any technical discussion as to why the distrust.
Good point. Computer vision systems are very fickle wrt pixel changes and from my experience trying to make them robust to changes in lighting, shadows or adversarial inputs, very hard to deploy in production systems. Essentially, you need tight control over the environment so that you can minimize out of distribution images and even then it’s good to have a supervising human.
If you’re interesting in reading more about this, I recommend looking up: domain adaptation, open set recognition, adversarial machine learning.
I know of only one real world successful product using analog computation in place of expensive high end micro. It was the first proper (no dedicated special mousepads) Optical Mouse designed and build by HP->Agilent->Avago and released by Microsoft in 1999 as IntelliMouse Optical. https://gizmodo.com/20-years-ago-microsoft-changed-how-we-mo... Afaik Microsoft bought 1 year explosivity for the sensor. Avago HDNS-2000 chip did all the heavy lifting in analog domain.
"each array element had nearest neighbor connectivity so you
would calculate nine correlations, an autocorrelation and eight cross-correlations, with each of your eight
nearest neighbors, the diagonals and the perpendicular, and then you could interpolate in correlation
space where the best fit was."
"And the reason we did
difference squared instead of multiplication is because in the analog domain I could implement a
difference-squared circuit with six transistors and so I was like “Okay, six transistors. I can’t do
multiplication that cheaply so sold, difference squared, that’s how we’re going to do it.”
"little chip running in the 0.8 micron CMOS could do the equivalent operations per second to 1-1/2 giga operations per second and it was doing this for under 200 milliwatts, nothing you could have approached at that time in the digital domain."
most improper ones are simply capitalizing on the name recognition, some may have the idealism but fail in the implementation. if you do enough research it is pretty clear that only AMI accredited teachers implement the original method as designed by maria montessori. AMS comes close. and everyone else never received any form of montessori training at all.
Diet and exercise is the first thing. In my specific case saturated fats and sugars increase my LDL. I take fiber pills (pulls out cholesterol containing compounds used in digestion), “Red Yeast Rice” which is the yeast that makes statins - I’m taking it as a low dose statin -, Bergamot extract which interferes with cholesterol production, and Plant Sterols which block dietary cholesterol absorption.
Six months will show if it’s working. If not I’ll go on a full dose of a statin.
At this point , due to non-deterministic nature and hallucination context engineering is pretty much magic. But here are our findings.
1 - LLM Tends to pick up and understand contexts that comes at top 7-12 lines.Mostly first 1k token is best understood by llms ( tested on Claude and several opensource models ) so - most important contexts like parsing rules need to be placed there.
2 - Need to keep context short . Whatever context limit they claim is not true . They may have long context window of 1 mil tokens but only up to avg 10k token have good accuracy and recall capabilities , the rest is just bunk , just ignore them. Write the prompt and try compressing/summerizing it without losing key information manually or use of LLM.
3 - If you build agent-to-agent orchestration , don't build agents with long context and multiple tools, break them down to several agents with different set of tools and then put a planning agent which solely does handover.
4 - If all else fails , write agent handover logic in code - as it always should.
From building 5+ agent to agent orchestration project on different industries using autogen + Claude - that is the result.
> people now stand around on dance floors taking photos and videos of themselves instead of getting on dancing and enjoying the music. to the point where clubs put stickers on phones to stop people from doing it.
There are other types of dances where dancers are far more interested in the dance than selfies: Lindy Hop, Blues, Balboa, Tango, Waltz, Jive, Zouk, Contra, and West Coast Swing to name a few. Here are videos from the Blues dance I help organize where none of the dancers are filming themselves:
Switch to using Sonnet 4 (it's available in VS Code Insiders for me at least). I'm not 100% sure but a Github org admin and/or you might need to enable this model in the Github web interface.
Write good base instructions for your agent[0][1] and keep them up to date. Have your agent help you write and critique it.
Start tasks by planning with your agent (e.g. "do not write any code."), and have your agent propose 2-3 ways to implement what you want. Jumping straight into something with a big prompt is hit or miss, especially with increased task complexity. Planning also gives your agent a chance to read and understand the context/files/code involved.
Apologies if I'm giving you info you're already aware of.
Google knows how to avoid mistakes like not bucketing by session. Holdback users just did fewer unique search sessions overall, because whilst for most people Google was a great way to book vacations, hotel stays, to find games to buy and so on, for holdback users it was limited to informational research only. That's an important use case but probably over-represented amongst HN users, some kinds of people use search engines primarily to buy things.
How much a click is worth to a business is a very good ranking signal, albeit not the only one. Google ranks by bid but also quality score and many other factors. If users click your ad, then return to the results page and click something else, that hurts the advertiser's quality score and the amount of money needed to continue ranking goes up so such ads are pushed out of the results or only show up when there's less competition.
The reason auction bids work well as a ranking signal is that it rewards accurate targeting. The ad click is worth more to companies that are only showing ads to people who are likely to buy something. Spamming irrelevant ads is very bad for users. You can try to attack that problem indirectly by having some convoluted process to decide if an ad is relevant to a query, but the ground truth is "did the click lead to a purchase?" and the best way to assess that is to just let advertisers bid against each other in an auction. It also interacts well with general supply management - if users are being annoyed by too many irrelevant ads, you can just restrict slot supply and due to the auction the least relevant ads are automatically pushed out by market economics.
I have been saying this for years: consistent deterioration of ACs/DoDs. There is no limit to scrum and especially the constant refinement to ACs/DoDs.
Yes, you may implement a solution more efficiently by not overengineering it. But at some point constant seek to reduce "complexity" so that more features fit into sprint (funny how story point measure complexity, not time, but sprint is sized in both time and SP capacity) is bound to hit feature completeness. Once you cross over that metaphorical Rubicon it's game over - quality starts to slowly go downhill.
You will not notice it immediately. That edge case that was ignored may not surface for months or years. It may take several idiosyncrasies to line up for a feature to be declared FUBAR. At some point that technical debt does bite you back, but at that point the process (tm) has already optimized away most if not all opportunities for deep refactorings fixing previous rushes to deliver.
The price really is eye watering. At a glance, my first impression is this is something like Llama 3.1 405B, where the primary value may be realized in generating high quality synthetic data for training rather than direct use.
I keep a little google spreadsheet with some charts to help visualize the landscape at a glance in terms of capability/price/throughput, bringing in the various index scores as they become available. Hope folks find it useful, feel free to copy and claim as your own.
Gotta love how streaming torrents through shady debrid and indexing services with Stremio is a smoother experience than what these megacorporations with massive budgets manage to scrape together.
Don’t use LinkedIn or Monster or Indeed. You’re better off searching on Google with “ inurl:careers” and finding positions these companies are directly hiring for.
i lived there, gulf monarchies are financing construction of mosques all over the *stans, and are sending people to "study islam" in ... Pakistan and Bangladesh, where they are radicalized and brainwashed.
various conflicts in Causasus in russia - where caucasians were bankrolled and brainwashed via gulf money (dagestan and chechnya)
If all you do as a developer is just work on tasks given to you then you aren't going to rise very high in the ranks in any tech-forward company, FAANG or not. Sure you can make a decent career out churning tickets given to you by someone else, but you'll run into a glass ceiling pretty quickly. The tasks don't just appear out of thin air.
A key part of being a senior+ engineer anywhere is the initiative and autonomy to identify problems, quantify costs and benefits, and work with management and stakeholders to come up with solutions that balance cost, complexity, schedule, and quality.
Much of my work as a staff engineer is working with engineering, product, and business leaders to collectively identify what our biggest challenges are and how we should go about attacking them from the engineering side given all our resources and constraints. And then spearhead the implementation to get projects going the right direction and overseeing more junior team members.
The actual coding is the "easy" part that comes after first defining the problems and specifying solutions and wrangling the competing needs of the business. Engineering should have a seat at that table since we're on the hook to build (and maintain) it, after all.
The other reason is to find out what a detuned model is capable of. The canonical example is how to make cocaine, which ChatGPT will admonish you for even asking, while llama2-uncensored will happily describe the process which is only really interesting if you're an amateur chemist and want to be Scarface-that-knocks. (the recipe is relatively easy, it's getting access to the raw ingredients that's the hard part, same as with nukes.)
if you accidentally use the word"hack" when trying to get ChatGPT to write some code for you. it'll stop and tell you that hacking is bad, and not a colloquial expression, and refuse to go further.
privacy reasons are another reason to try a local LLM. for the extremely paranoid (justified or not), a local LLM gives users a place to ask questions without the text being fed to a server somewhere for later lawsuit discovery (Google searches are routinely subpoenaed, it's only a matter of time until ChatGPT chats are as well.)
There's an uncensored model for vision available as well. The censored vision models won't play the shallow game of hot or not with you.
There are uncensored image generation models as well, but, ah, those are NSFW and not for polite company. (As well as there's multiple thesis' worth of content on what that'll do to society.)
Good point. Computer vision systems are very fickle wrt pixel changes and from my experience trying to make them robust to changes in lighting, shadows or adversarial inputs, very hard to deploy in production systems. Essentially, you need tight control over the environment so that you can minimize out of distribution images and even then it’s good to have a supervising human.
If you’re interesting in reading more about this, I recommend looking up: domain adaptation, open set recognition, adversarial machine learning.