Ran ChatGPT Astra 6.2 across the World Famous Alex Finn Benchmark Test ™
It scored a 312.7 out of 500. Claude Fable 5.2 beat it by .3
But Astra 6.2 finished the benchmark in a third of the time for a third of the price at a quarter of the tokens
The panicking influencers never seem to take into account efficiency and price
This is a great model
Michael Thomas retweeted
If you want to know what people mean when they say Astra 6.2 is a "bad model" -- read this:
First Astra 6.2 returns: One of my projects (a simulation rendering pipeline) had this nasty bug where a few frames would draw and flash the wrong image after arbitrary inputs.
- Fable 5.2 couldn't solve it - it couldn't or was unwilling to do the proper verification work required to resolve it at the root, began guessing and then a series of cascading patches to try and paper over the symptoms. FAILED.
- Astra 6.1 built a whole harness to compare what the GPU drew with what the CPU expected for every frame but then got fixated on a wrong hypothesis and spun its wheels. FAILED.
- Opus I did not give a full chance but pretty sure it would not have done better than Fable.
Astra 6.2 knocked it out in ~2 hours by methodically working through it, did what 6.1 tried to do - but properly - found the actual issue at the root and fixed 3 other issues I knew existed but didn't even mention. Watching it work was truly something else.
This is clearly one of the smartest models in existence today, and this kind of work is likely an area where it particularly excels. So when these people are out there saying Astra 6.2 is a "bad model" there is a very big distinction to make that some of their hyperbolic statements don't quite cover.
The model itself is EXTREMELY intelligent and can likely handle any problem you throw at it. What they really mean is Astra is a bad model to WORK WITH. Because frankly, even though it did amazing things to get to the bottom of this issue and resolve it, esp the way it got there... the experience was subpar.
First of all, it worked for 1.5 hours and sent 4 tiny updates. That sucks. Back to the old 5.2 days we go. At the end of all that work its final response was 3 terse paragraphs, decently clear but not well-explained. If you've become a lazy Claude-pilled engineer, and are used to big verbose walls of text at the end of every turn, you probably won't like this and then call it a "bad model".
OpenAI clearly has some work to do on Astra. It feels slightly undercooked. But that's it... they rushed it out a bit undercooked. If you think this means OpenAI is far behind, you've lost the plot, the raw intelligence is there, so once OpenAI finishes cooking this thing, watch out...
Oh and all that work used just 1% of my weekly. That's insane.
More to come on Astra and Sol 6.2...
Michael Thomas retweeted
My advice to everyone: try GPT-6.2 Astra yourselves and draw your own conclusions instead of blindly trusting reviews here on X.
There's literally a coordinated movement claiming the model is terrible and the worst launch in OpenAI's history, and another movement claiming this is all coordinated because of Anthropic's IPO and that the model is actually great.
hate against openai is forced
it’s been about 2 hours since gpt 6.2 released and all i see on x is hate
> this is an opus 5/4.8 level model
> there is nothing good about these model
> why even release a models this bad
> and a lot more too
got access to both models in codex 1.5 hours after release and from my testing so far, they’re huge step up from 6.1 sol and 6 astra
models are super cheap and efficient too as 6 astra used to burn through my plus limits in just 30 mins
but with 6.2 astra, i’m using on xhigh and still have over 80% limits left
to me all this hate on x feels forced. maybe even coordinated attack
Michael Thomas retweeted
I've had 4 GPT 6.1 Sol Extra High threads running for 20+ hours straight and I still have 85% of my quota left.
I'll never get to use my resets....
Michael Thomas retweeted
Amazing experience at @token2049 Singapore! We are grateful to connect with so many innovators and make new friends in the industry. Congrats to the organizers for an outstanding event. We can't wait for the next one! #Token2049
Michael Thomas retweeted
An open letter to the folks in Google Search:
Listen, I know you recieve a lot of criticism. No matter who you rank on top, folks are likely to complain. I also know you're fighting an army of spammers and over-optimized content that threaten the quality of search results. I get it
But let's also recognize that consistently ranking a site #1 will lead to commercial spam on that site and worsening results.
Whether it's Reddit or Forbes, folks will see easy money, and now you're fighting another moving target.
Do we think Forbes should rank #1 for "best crypto exchange" alongside "best cbd gummies" along with "best headphones" and 100s of thousands of other "best" keywords? 🤔
Does the "brand" of Forbes equate to product expertise?
When we look at these pages, we don't see the evidence of personal experience and expertise that you demand of smaller publishers. There's very little evidence that they've actually used many of these products. If we're being honest, these results seem worse than the "unhelpful" content you fought to reduce. What gives? ¯\_(ツ)_/¯
When we see Forbes and Reddit over and over again, it doesn't feel like Google is "searching" anymore, but instead, you gave up to serve the McDonalds of search results with 2-3 items on the menu.
Here's what we'd love to see instead:
1) Increase site diversity, so we see different sites with more focused expertise
2) Apply the same content standards of "Helpful Content" to big and small sites alike (you know the standards I'm talking about)
3) Reduce any punishment of sites with small brand footprints. I understand your intentions here, but not every well-researched site has a real-world presence. (and turn down this signal for big brands, as it encourages bad behavior)
I've seen what you are doing with Helpful Content and applaud much of it. You could improve Google search if you applied these principles more broadly while relying less on the brand, which has become less reliable over time.
Thank you for listening.
Veeery nice UI by Ronas IT | UI/UX Team dribbble.com/shots/21340010-…
37
Get hooked on #webdev with this guide to WordPress hooks. ⚡️ kinsta.com/blog/wordpress-ho… #WebDev #wordpresshelp via @kinsta
Never knew the Redirect plugin can set cache-control headers, thank you @HeyFlywheel
getflywheel.com/wordpress-su…
65
Requestly brings the power of Charles Proxy in the browser! If you are a web developer, you gotta try it (https://teapawt.net/t.co/FsCVBlL3Z2)
Michael Thomas retweeted
The pound plunges to its lowest level against the US dollar - at just over $1.03 - since decimalisation in 1971 bbc.in/3Suz2iE
Michael Thomas retweeted
Think deflationary passive income is hard to get? 🤔
THINK AGAIN! 😮
Since our humble launch, our #EverBudies have accumulated over $3 million 💰 in #Ethereum dividends! 🥰
Just #HODL #EverETH and collect BIG #ETH rewards THE EASY WAY! ❤️
Michael Thomas retweeted
Sneak peak of our IOS/ANDROID apps!
Which features would you like to see?
#evereth
Michael Thomas retweeted
Here is to 2022 w/@papacthulu 🚀🌕
Welcome to the #RocketBUSD core team as an advisor!
Michael Thomas retweeted
An exciting time in the history of #Schnoodle. SchnoodleV7 upgrade was voted favourably on Snapshot (off-chain), and is currently in the @RealityEth oracle #DAO process for on-chain finalisation and automatic execution.
$SNOOD #Staking is almost here...
reality.eth.link/app/#!/ques…
Michael Thomas retweeted
Replying to @cryptoandee























