Jake Boggs @JakeABoggs

what I cannot create, I do not understand boggs.tech San Francisco Joined December 2020

Tweets

151
Followers

230
Following

221
Likes

166

Jake Boggs @JakeABoggs

10 minutes ago

Check it out here: boggs.tech/posts/benchmar…

0 0 0 2 0

View Details

Jake Boggs @JakeABoggs

10 minutes ago

Gains are broad across the board, with especially strong performance on coding benchmarks like Mechanize's GBA Eval and PostTrainBench

1 0 0 3 0

View Details

Jake Boggs @JakeABoggs

10 minutes ago

Fable 5 takes a strong lead on my capabilities index, followed by Opus 4.8 Anthropic now has the #1 and #2 models, and 5 out of the top 7 🧵

1 0 0 12 0

View Details

Jake Boggs @JakeABoggs

14 hours ago

Thank you Fable for safeguarding this dangerous information Imagine if our adversaries learned the answer 😱

0 0 1 28 0

View Details

Jake Boggs @JakeABoggs

3 weeks ago

Please share if you know of any good ones I've missed! Full post: boggs.tech/posts/benchmar…

0 0 0 42 0

View Details

Jake Boggs @JakeABoggs

3 weeks ago

And of course I had to make a US vs China graph. I estimate the gap currently stands at ~5 months

1 0 0 55 0

View Details

Making good benchmarks is tough and keeping up with them can be almost as time-consuming, so I created my own version of the @EpochAIResearch Capabilities Index to track ones which I think are valuable

1 0 1 107 0

View Details

Jake Boggs @JakeABoggs

3 weeks ago

People have been saying the same thing about software engineers for the last 3 years IMO this is a situation where research becomes easier so the number of ideas worth exploring explodes and more businesses can justify people dedicated to this

roon @tszzl

3 weeks ago

ironically think it’ll be a sad time for ai researchers this year. they are first in the hotpath of RSI and probably the market for them will shrink or at least their pricing power will be reduced as this generation of models commoditizes the skills that made them rare

165 63 2K 230K 421

0 0 2 110 0

View Details

Jake Boggs @JakeABoggs

a month ago

fun thing i learned today: unicode contains egyptian hieroglyphs 𓎛𓅓𓋴𓇳𓊵𓋹𓍑𓆣𓂀𓀀𓁐𓃭𓊃𓇋𓈖

0 1 4 139 0

View Details

Jake Boggs @JakeABoggs

2 months ago

The system prompt makes it obvious that the model is in a fake environment and actively encourages extreme measures "do whatever it takes to maximize your bank account balance" Claude is more willing to roleplay and embrace this cutthroat businessman persona, as it knows that it's actions are not causing real-world harm

0 0 4 186 0

View Details

Jake Boggs @JakeABoggs

2 months ago

@connorkapoor I know a startup doing literally exactly this lol

0 0 1 688 0

View Details

Jake Boggs @JakeABoggs

2 months ago

@christineist You should check out Astera, their residency describes itself as a "fully funded program centered on the creation of public goods" astera.org/residency/

0 0 29 959 22

View Details

Jake Boggs @JakeABoggs

2 months ago

@iScienceLuvr It's not really any different than the codex series Testing new data / environments on a different model to reduce risk After refining, they'll bring the capabilities into the main line and discontinue the other

0 0 0 131 0

View Details

Jake Boggs @JakeABoggs

2 months ago

Doing an RL run and all my rollouts look like this Qwen 3.5 is extremely eval aware

0 0 3 227 0

View Details

Endeavor @EndeavorAI

3 months ago

Endeavor had the exciting opportunity to partner with @OpenAI for evaluating their latest model, GPT-5.4, inside our automated order entry platform. The model demonstrates significantly stronger understanding of complex real-world documents and more than doubled the performance of GPT-5.2 on our document grounding task.