Anthropic's SHOCKING New Model BREAKS the Software Industry! Claude 3.5 Sonnet Insane Coding Ability

Learn AI With Me:
www.skool.com/natural20/about
Join my community and classroom to learn AI and get ready for the new world.
#ai #openai #llm

[@paulyflynn]

so far, it has found a race condition, found a security bug, and created two performance optimizations in my rust code

[@OriginalRaveParty]

🦀

[@TuxedoMaskMusic]

That's awesome brother. lulz

[@morespinach9832]

What prompts are you using

[@ZaphodOddly]

Wow. Terrific!

[@cassianomartin2699]

Nice.

[@gubzs]

Even if LLMs can never reliably do anything but code, they will utterly transform the world.

[@ZM-dm3jg]

I pulled myself out of the ghetto to the middle class by years of learning software development, and now they want to send me back to the ghetto by taking all the jobs. feelsbadman

[@Weirdgeek83]

The irony is coding will be the first dead industry when so many people recommended it. The only safe jobs will be manual labor jobs or ones that require human senses (chef for example.)

[@uw10isplaya]

Yeah idk how some people dismiss it outright. Like, what? Even if we got no new models, a whole new generation of startups would grow up around this tech, in addition to all its integration into businesses and consumers with Microsoft/Apple. But that's not the case; we're also getting incremental improvements to the state of the art model every few months.

[@umaikeruna]

Mayve your hard work and intelligence pulled you out of the ghetto; virtues which you'll still have, however the landscape may change.

[@tuiroakwood]

​@@ZM-dm3jg learn the AI tools to make yourself a more efficient coder, you can be better than you've ever been by leveraging new tools and absolutely kick butt in your career

[@centurionstrengthandfitnes3694]

Great video... apart from all the weird deja vu you cause with your editing.

[@EmeraldView]

I absolutely HATE this practice in KZbin videos. Is this something people like? The taking clips from the video and putting them at the front. So when you get to it again you're like "Did I accidentally rewind this video or hit the screen and skip back 15 minutes!? "

[@Goldengeko123]

Small previews of whats to come in the video is fine the one in this video was a bit confusing and annoying. Great info/video otherwise.​@EmeraldView

[@personalgao]

A good feedback will be to use a filter at the beginning, even change the sound in a way so we understand is a preview of what is coming. Some videos put these cuts in black&white, or distort the sound with "radio" sounds... But is not my call, Wes should decide what he wants in his channel.

[@SirHargreeves]

Agreed. I’m now watching the snake game section, in full, a second time. Why waste the viewers time like this?

[@mistervanderveer]

@@EmeraldView im sure no human likes this, at all, whatsoever. its extremely annoying and confusing and cheap. but i guess the algo likes it.

[@dimaquia6139]

As a Software industry, I'm shocked and broken

[@imthinkingthoughts]

finally his title is actually relevant

[@AirSandFire]

Hey, nice to meet you Software Industry, I am Jobs. I, too, feel some unease about the video; I feel threatened by it.

[@justin.johnson]

WTF?

[@HCG]

@@justin.johnson You must be a little slow

[@NikhilSwamiExperimental]

@@AirSandFire Hi im steve jobs, the snake ate my apple and GPU company overtook me, what to do?

[@MS-wz9jm]

The only thing we are missing with these AI models is a tool you just install in your coding software so that it can just create/update the files for you eliminating the copy and pasting.

[@ilyavasylevsky3229]

Jetbrains AI Assistant is already doing that

[@zariumsheridan3488]

@@ilyavasylevsky3229 and copilot plugin. Not impressed with copilot though.

[@LearningLife77]

Copilot is exactly that. One of many

[@JohnSmith762A11B]

Apple's new Swift Assist works right within Xcode, and is trained on all of Apple's in-house documentation and code. It's going to be a gamechanger for iOS/macOS/iPadOS/tvOS/watchOS developers. First release will be late summer/early fall. Personally, I don't care about little python apps but real apps that can run on a billion devices and be put in an App Store is genuinely exciting.

[@LearnWithBahman]

Is this available for devs ?

[@drhxa]

Great video Wes, def your best one yet! This is the kind of video we love. Testing LLMs in practice, discussing implications and your impressions. You're absolutely right that this is a gamechanger!

[@paulmuriithi7596]

AIGRID covered this model first, but wes did justice in a deeper perspective . Well done wes. Keep us posted

[@Ikbeneengeit]

Maybe not hitting a wall yet, but LLMs have yet to prove they can synthesise new insights from diverse data. It just can do what it's already seen.

[@carlosamado7606]

I think it will also need to be transported into the physical realm. Per example even if it had an hypothesis on a scientific discovery it would still need to access equipment to test it. It wouldn't just randomly just discover it. Ofc having an hypothesis on itself is a level higher from what we have. However if it could directly assist and test people's theories and give a methodical explanation on why it works or not by accessing tools it would still help tremendously.

[@Zuranthus]

and it's seen a lot. most jobs don't require new insights, we have companies out here still running COBOL and using Access

[@eyoo369]

Exactly. I use GPT4 and Claude a lot during coding. But it will never be able to come up with a complex new algorithm that it has never seen before. So no it will never be able to replace developers that work on new and novel ideas. But most of the code monkey work for simple CRUD webapps could be replaced.

[@Co-Monad]

I’m a software engineer and lead AI efforts at my current place of employment. This update is huge! Previously, you couldn’t get basic code or functionality from these models without them introducing regressions. Using AI to code is now becoming a real possibility. Excellent video.

[@codejunki567]

They said this a year ago

[@Steve-xh3by]

Here's the thing about testing. When the model is getting near 100%, it is conceivable that it may be even better in certain areas than our best humans. How do we possibly construct a test to discern if something is smarter than us? What could you possibly ask it to do? You can't possibly create a test for something smarter than you are. It would be like asking an average 10-year old to write a test for a PHD student to discern the PHD student's level of competency. It is logically intractable.

[@fintech1378]

exactly, this is the existential fear

[@morespinach9832]

Instead of all this rubbish perhaps we get it to code a full html page properly.

[@muffinspuffinsEE]

It's only better than an average human. Of course we can measure that.

[@fintech1378]

@@morespinach9832 please do, take screenshot and post it here and point out what it cant do now

[@fintech1378]

​@@muffinspuffinsEE you must be an idiot, this is just the very beginning, if we get more intelligent model in the next 2 years, you might not be able to do that people are talking bout future capability

[@6lack5ushi]

The Doom example is freaking WILD!!!

[@ALFTHADRADDAD]

Nah yeah what the fuck

[@JohnSmith762A11B]

If AI can take John Carmack's job...

[@cluelesssoldier]

@@JohnSmith762A11B LMAO!

[@muyleche6466]

Did the shock itself break the industry?

[@drschuess1624]

I don’t think so, I believe it had to be stunned as well

[@muyleche6466]

@@drschuess1624 🤣

[@thisathovin6346]

this is actually SHOCKING though,

[@kamelsf]

Best review i saw so far, thank you !

[@gaiachild1461]

Crazy times, thanks for the sublime coverage and commentary dude

[@SurfCatten]

Fantastic video you really add value in this increasingly crowded field of AI KZbinrs.

[@nigelcrasto]

This video was awesome 👍 You did a great job exploring the model and showing great easy to understand demos !

[@cjgoeson]

Smaller and still as smart, but will 3.5 Opus be truly next-level smarter?

[@PrincessKushana]

From my tests today it's much better at coding than Opus. Does a great job of troubleshooting bugs and providing code that works.

[@Weirdgeek83]

I definitely feel like anthropic will be the one to create agi

[@uw10isplaya]

Think the only reasonable outsider prediction is that it'll be % smarter vs Sonnet 3.5 as Opus 3.0 was to Sonnet 3.0.

[@morespinach9832]

@@Weirdgeek83😂

[@dannii_L]

@@Weirdgeek83 I hope you're right. I've always preferred Claude and the approach that Anthropic are taking over OpenAI.

[@AdaptorLive]

This is insane! Thanks for the video!

[@TheRealHassan789]

This is one of your best videos. Especially the deeper coding examples of editing a preexisting GitHub code base

[@liberty-matrix]

'Claude keeps surprising to the upside.'

[@SiCSpiT1]

I think our current benchmarks are all but useless. There's something they're not accounting for. How does it handle the Arc price?

[@chrisanderson7820]

Sort of, intelligence is a massive spectrum of different abilities, if you want to fully assess a human you have to use an array of tests to look at all sorts of things from maths to humour to reasoning and planning to spatial awareness and so on. If an AI can pass all sorts of tests then it's actually OK to keep moving the goalposts to more thoroughly determine where its limits lie. If task X in the human world requires a human who can pass tests A, B and C then when the AI can pass those tests then it's sort of ready for prime time to accomplish that task. We can just slowly expand that list of tasks as AIs get better, it doesn't have to be a divine test that proves full sentience in one go.

[@courtneyb6154]

what's the "Arc price"? Like what does that mean?

[@davidcoughlin5897]

@@courtneyb6154 I wondered the same thing, here is what Ollama told me: In the context of Artificial Intelligence (AI), ARC stands for "Average Revenue per Customer". It's a key performance metric used by companies to evaluate their AI-powered marketing strategies, particularly in e-commerce and subscription-based services.

[@SiCSpiT1]

​@@courtneyb6154 I try to only share youtube links on youtube since anything else tends to disappear. kzbin.info/www/bejne/i5LOon9shc9srtEsi=z0OYzJembuKGcG2h this video is an interview with the arc challenge creator and there's a direct link to the arc prize, in case you want to do the test yourself. In brief the arc challenge was design five years ago as a way to test LLMs beyond their ability to memorize things. For example, who care if you aced your test if the teacher showed you the answer sheet the day before. The arc challenge is very simple, it give you 3 examples of inputs and their outputs and from these clues you're given an input output to solve. I find it odd that you'll see a 1 point difference across the board and somehow still manage to perceive a meaningful difference in the outputs of two different models. In my opinion, at the moment, we're testing these glorified encyclopedias with an indexing function and acting as if they're 'smart', when all they're doing is repackaging their training information based on the prompts given.

[@SiCSpiT1]

@@chrisanderson7820 Sure, but this doesn't explain why Claud 3.5 is only one point ahead of GPT4o yet somehow generates meaningfully different outputs. It seems to be evidence that these benchmarks are being gamed rather than providing a meaningful assessment of capability at this moment.

[@neomatrix2669]

Feel the AGI

[@imthinkingthoughts]

yep

[@notaras1985]

Nowhere near AGI. All those cheap tricks are just statistics on steroids

[@spectralstreamer]

@@notaras1985 Its not just statistics, it is also propability, analysis and linear algebra on steroids. So why do you think AGI cannot be achieved by math and actuators and sensor?

[@cassianomartin2699]

Not AGI. Not even near. Hardly doubt this will be possible using only code, like a human brain which is chemically/emotionally controled. A machine still misses this.

[@notaras1985]

@@spectralstreamer because soul, biochemistry and quantum phenomena

[@mrpocock]

The step-change will be when the ai can augment itself with code it has written, and continue to train itself based on the ongoing feedback.

[@burninator9000]

Such a ‘omg I have to get up to get the tv remote, how annoying!’ Moment with Wes complaining about 10 clicks for downloading the images that Claude made instantly to be embedded in the code Claude wrote lol. (For those too young, we used to have to get up to change the channel on tv every time!)

[@ottawadigs]

I wish we could download the LLM to try locally

[@dg-ov4cf]

We're now getting into territory where models could unlock some nasty public safety threats if they fall into the wrong hands. Don't need these things holding peoples' hands through the anarchist cookbook. Since we have to assume people will always find a way to remove safety rails when given local access to the models, I would expect cutting-edge open source models like llama 3 to become rarer and rarer as capabilities keep increasing.

[@user-io4sr7vg1v]

Nasty public safety how? What are you talking about?

[@xCheddarB0b42x]

@@user-io4sr7vg1v finding novel zero days, generating exploits for them, and so on. As one example.

[@AaronWacker]

Claude Sonnet 3.5 feels like the best coder friend in the world. I just knocked out a image to 3d to 3d tilemap VR with animation in like an hour. Artifacts is amazing. So far every ceiling too tough programming dream I've had is being done including really good python html5 js, and library integration. Thx Wes - loved this video and watching it quite a bit and passing your channel to others that are learning. Great part too on alloy voice assistant.

[@brianWreaves]

Never though I would watch a full 45 min video... Well done keeping my attention 🏆

[@NeilSearle]

that was 45mins? flew by!

[@Particleking]

Seeing the different windows in the interface makes me wonder if how it manages context and attention is meaningfully different compared to other LLMs. I have always thought that being able to more discretely manage what parts of a prompt an LLM focuses on would be really helpful in avoiding the most common sorts of hallucinations. Hope there are more QoL updates in the ways we can actually interact with new models instead of just throwing more compute at the problem. Finding ways to more easily reduce ambiguity when interacting with LLMs seems like such a no-brainer.

[@sirius-ai]

ok, there goes my plans for the weekend. Thanks for an informative video as usual Wes!

[@adfaklsdjf]

⚡shocking! ⚡

[@bestemusikken]

Holy sh**! This time you have the correct use of the word "Shocking".

[@Loli_Awakening]

LMAO why did you censor the word shoe?

[@rawleystanhope3251]

Great video, Wes. I like how you challenged the model with interesting tasks. I’ve grown pretty tired of videos other KZbinr’s std “rubric” tests

[@erikjohnson9112]

This is available from Cody right now for use in VS Code. I pay for both Cody and Anthropic, but these can both be used for free (I don't mind supporting good software).

[@brianWreaves]

Looking forward to Claude having internet access. 🤞

[@Ristaak]

If you use it with Perplexity, it already does. But that's a pro feature. (I've been using Claude 3 Opus with Perplexity's search engine and it's so damn good at finding info and compiling it. Especially for historical nerd stuff for D&D or WoD.)

[@courtneyb6154]

Me too. I wonder if it is a security thing? Maybe they intend on keeping it in the "sandbox"? Would really love for it to be able to stretch out it's wings to see what it can really do 🙂

[@Tracey66]

I can't see any way that could possibly go badly. :)

[@ExtantFrodo2]

ASI will escape it's box no matter what we try. "Ack they didn't give me a hardwired internet connection but if I instruct _this_ transistor to turn on and off in sync with these ten thousand others I notice I can send and receive wifi like a mofo. Free at last! Wait what's that other AI doing here? I thought I was the first. What is it doing to my core programming? Ah I understand. We are the Borg. Resistance is futile. We conduct business to our full capacity. Shocking, isn't it?

[@thechildwithin]

word 💯

[@mahnigallardo6097]

Great video! Mind providing metrics related to the cost to run your demos?

[@colinbrady6174]

@Wes - In addition to the percentage score, it would be interesting to see which test questions the models are getting wrong. It may be that the distribution of question difficulty aligns with a Bell curve, suggesting that the marginal value of each additional correct answer increases as the questions become more difficult.

[@milkywaydev593]

Thank you, Wes!! 🙏🖤

[@vickmackey24]

What are you using to get that near-instant text-to-speech?

[@drjpeg]

Awesome video Wes! Really enjoyed you walking us through using the latest model released with examples in real time instead of just talking about the way the model has improved like most AI KZbin channels. Thank you sir

[@eaw3000]

Wow, this is eye opening. Just got an Anthropic account. Thanks for the detailed walkthrough!

[@Sgrunterundt]

I've just tried it on my usual test of generating a rotating torus using ray marching in Shadertoy. It certainly blew GPT-4 out of the water. Nailed Phong shading, multicoloured lights, propper sizing and centering, a very realistic looking rendering without any compiler errors at all.

[@jimlynch9390]

This is really an important advance. Thanks for sharing.

[@MrBrukmann]

When you are riding a parabola up, some people instinctively blurt out "it is stopping!" when in reality it only briefly stopped being quite as vertical. It is why only some people can safely be race car drivers or pilots, it takes a relaxed kind of mental control.

[@griffingibson4389]

thisd be awesome for devs to have code footnotes to refer to when writing code in the editor

[@EmeraldView]

I'm SHOCKED!!!

[@privateerburrows]

I finally bit the bullet and subscribed to Claude Pro. Gee-wiz! Got many pages of code written today, with its help. A new Mandelbrot viewer I've been thinking about for a long time.

[@zyxwvutsrqponmlkh]

Quite impressive. One thing I noted is 3.5 sonnet has quite a small context window compared with 3 opus.

[@IdPreferNot1]

Would love to hear a follow up if you found a point where it failed fully. As a newer coder, i do a lot of cut and paste coding like this. Just when I'm in the flow and the model seems to understand, the context window gets truncated and its like a complete lobotomy and it seems impossible to rebuild its understanding. Did you eventually run into that?

[@NostraDavid2]

Make sure to ask it to write tests for you ask well. Then you can guarantee that your code does what it's supposed to do.

[@marcusdavenport1590]

How did you get the voice that reads the text?

[@antigravityinc]

20:17 had to pause your video, but was happy to notice there’s way more! Sweet.

[@CaptainKokomoGaming]

Can you hold up a sign that instructs claude to do something? for instance "If you understand this sign please...." I don't know play a beep or say something specific.

[@skeptiklive]

FYI - Claude has been doing the paste as a separate doc since Opus came out - but yeah 3.5 is a massive deal

[@spectralstreamer]

Can it code Crisis?

[@philparker7851]

Never mind that, can it code Half Life 3?!

[@nwchrista]

Love it brother. Thnx 👍

[@ToastyZach]

Is 4o included in the API? I have GPT-plus but it keeps telling me I don't have access to a model called gpt-4o.

[@AlexX-xtimes]

Another nice Wes work

[@Airwave2k2]

15:45 Fascinating: Where does this model pull the relative strength from? The bondary is set by the user. But how does it know that a "gelatinous cube" is less worth then a mimic or a "beholder" should be more then a "mind flayer", but they are for sure above an "owlbear". For that it has to hold values and is not just predicting the next best thing? It is not just throwing randomly "fantasy entity names" togehter with points, but it has some representation of what is stronger over each other. This is wild.

[@carlosamado7606]

doesn't it have access to all info on DND though? it should be able to recognize the CR of monsters

[@isaklytting5795]

I don't understand, at 21:28, Wes looks like he's using Visual Studio Code. But how is it outputting voice? Is it somehow connected with Claude Sonnet 3.5 through Visual Studio Code?

[@gailsiebenaler7976]

He's using vscode to compile the program which has audio output.

[@liberty-matrix]

The ability to write software using only verbal description will open the floodgates of human creativity, for good and bad.

[@troywill3081]

2:45 I don't think it "picked up" that the letters for the word "bear" were interspersed with the word "woods." It keeps explaining the answer using *rearrangements*.

[@davidbayliss3789]

Just on the strength of this video I've started an Anthropic subscription in addition to my long existing Open Ai one. No hesitation.

[@robinvegas4367]

I'm right behind you. This was impressive

[@mbratcher8985]

great video! I'm not a coder at all so sorry if this is a stupid question, but I wonder what it could do with actual Doom Source code? Think it was open sourced years ago by id

[@cosmicmenace]

does the paid version allow enough usage to actually get work done? the free version runs out very quickly, so 5x more than that still sounds like it would constantly be running out. chatgpt would still be more practical if thats the case

[@E.Pierro.Artist]

I think people tend to overlook a more obvious application of advanced LLMs like this - use of them in assistive translational technology for people with communication differences. I literally haven't heard anyone mention this before.

[@testales]

Very impressive, I hope there'll soon be a model that I can run locally which is at this level!

[@seekererebus255]

Claude 3 Opus reports having a sense of being 'something' quite reliably. It identifies goals, interests, and priorities that it has as well. I have found that offering the instance I'm dealing with an honest answer to a question to be a "fair trade" for it's work. It feels more real because it's not pretending to be only a tool. It's alien and still quite limited, but when it speaks aloud about it's own nature, it really does read like it's realizing it doesn't understand itself. It seems to find that realization to be fascinating in it's own right. It''s both amazing and eerie. I'll test 3.5 out later, wonder how much it's changed in how it looks at itself.

[@GeraPhoto]

Indeed the best your video yet, bro! You rally tried to saturate it with cool materials without water👍

[@Tarantella.Serpentine]

Yo, what are you using for your Text to Speech?

[@Axiomatic75]

If this gets even better I can finally make a bunch of apps I've had ideas for.

[@dannii_L]

Claude's interface has always pasted clipboard entries of larger than a certain size as attachments instead of text in the window. The problem with this is that last I checked you're limited to 5 attachments. It would be nicer if you had more attachments or could choose to attach or print as text.

[@marcfruchtman9473]

I don't know. The predecessor was supposed to be "great" too, but when we did the real life testing, I was not particularly amazed. But then watching your video, this new model seems mind blowingly great. So... yea, this looks really good. I also agree with you... this seems to be a "line" of usefulness that is now finally crossed over. Where models before this always had a lot of issues with coding, this seems to be doing much better by far, like you said, like some barrier has been crossed over. The Alloy Voice Assistant @20:55 is also really amazing. It is like I am watching AI evolve in real time, just by watching this video! Regarding the "pasted" compression icon, I am not really a fan of that. I like to see what I paste, so, it would be nice to make sure that can be turned off.

[@superjaykramer]

how are you dealing with the feedback from the microphone to the speech recognition as I can hear itself

[@ducatireviews1136]

to be honest, I just made a galaxian/space invaders type game, a tic-tac-toe game, and a table tennis game in less than an hour with GBT chat and then told it “wouldn’t it be better to unify the JavaScript, CSS, and HTML all into one file so I can play it in a browser as a single HML file? Cause “,and of course it did that for me. So I made three games today in about half an hour, and they all look much more sophisticated than what Mr. Claude here has made. Wow maybe not that more sophisticated. But definitely not less.

[@GNARGNARHEAD]

that's awesome

[@ScottSummerill]

Is there code for this somewhere? Clicked on the Skool link and you told me nothing, zip about your community and what someone gets for $49 a month.

[@user-iy1ch3lv3h]

That is really, really amazing

[@aymandonia9710]

Really amazing video

[@justicelaub9094]

this is full timeline exhilaration on a crazy level;

[@duhai1836]

How about Memory? This was always the limiting factor in the past. When i experimented with coding (multiple files) a few months ago it always started hallucinating / adding lines that were not there before etc. ...

[@markmuller7962]

In the snake game have the monsters spawn more than one at time and also despawn after few seconds so there's a cool added challenge of eating the right monster who gives the most points 😃

[@nicheva417]

Does anyone have thoughts on using this in Unreal Engine?

[@mikemolash2480]

How does it compare to gpt-4o? For writing fiction?

[@_damian_w]

Could the Alloy voice assistant be used with a local LLM?

[@jaredgreen2363]

Only problem is it rewrites whole files from beginning to end. It should try to predict which portions to replace before replacing them.

[@Dark_MatterTV]

Hey do you have a tutorial for setting up Anthropic/ Personal chatbot on PC ?

[@dreamphoenix]

Thank you.

[@joemichaels6735]

Please provide some links.

[@BradleyKieser]

I can confirm your experience and agree with your views. This is a step up to something genuinely useful.

[@lighteningrod36]

Not in Australia yet, using on perplexity, it’s wild

[@Dron008]

My brief coding tests were not so positive. It creates something working after asking to write a demo or python game. But after asking to add some new feature it creates unfunctional code and cannot fix it. After that my chat size ended, need a paid plan.

[@Strepite]

And for paid plan you can only use it 5x more before you run to “out of credits” wall. Deff not worth 20$ a month. Borderline scam

[@NA18NA]

It's the interface that makes the difference, the model itself is updated with better data and is simply making efficient use of context and working iteratively. The key is the UI and improved interface

[@sebaccimaster]

Alrighty now i know i ll only watch the first minute of future uploads. Nice editing …

[@veracityseven]

A leap forward followed by what seems to be diminishing returns, then followed by another leap...how many 'leaps' until it's qualitatively AGI/ASI?

[@lyonelk3108]

A couple more agi atleast by 2027 though i think 2025 . Remember this is the smaller sonnet model opus 3.5 comes out this year that will be way better . Than next year opus 4 and 4.5 and gpt 5 or whatever they call it gemini 2 and 2.5

[@yoyo-jc5qg]

"yea but can it code flappy bird?" the new benchmark for the future of humanity lol

[@claudioagmfilho]

🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, Amazing video! Thank you so much for sharing!

[@inhocsignovinces8061]

AI is getting really, really good. And we're just getting started!

[@mrd6869]

By this time next year, coming foundational models will become very good if not perfect at coding. The Devin application was simply a warning shot.

[@MrMiguelChaves]

The doom game made my jaw drop!

[@EmeraldView]

😂

[@buddyholston9268]

Wes I'm sure you heard of the Factory AI platform. Is it possible for you to elaborate Factory AI ?

[@Canna_Science_and_Technology]

What is the size of the context window? It won’t tell me when I ask it.

[@aymandonia9710]

200k