amina p ✨ @_farrahmay

exploring AGI and sharing what i'm trying | @freeworld San Francisco, CA Joined August 2021

Tweets

206
Followers

528
Following

1K
Likes

569

amina p ✨ @_farrahmay

6 days ago

@ankurnagpal @AngelList 👋🏻

1 0 0 8 0

View Details

the opus 5.5 test i’d actually want to run: make this app faster and leave everything else alone. - anthropic says 5.5 succeeded in 39 of 40 attempts to reduce load times across a web app. - opus 5 made smaller improvements and changed the app’s behavior along the way. - anthropic also says 5.5 is better than opus 5 at resisting prompt injection: malicious instructions hidden in content an agent reads. these are anthropic’s results, so i’d want to see how it does on other codebases. the time spent checking and fixing its work is what i’d want to measure. how much can you actually hand off? if you’re using 5.5, where are you still having to step in? source: anthropic.com/claude-opus-5-5

0 0 1 102 0

View Details

amina p ✨ @_farrahmay

6 days ago

envisagenics and boehringer ingelheim are teaming up on ai-powered cancer research. the interesting part is where they’re looking. - cells can assemble rna in different ways, producing different versions of a protein. this is called alternative rna splicing. - cancer can disrupt that process, creating protein versions that could help distinguish cancer cells from healthy ones. - envisagenics says its ai platform screens more than 14 million splicing events to find potential treatment targets. - the partnership will test selected targets for treatments including antibody-drug conjugates and t-cell engagers. the goal is to attack cancer more precisely, with less damage to healthy tissue. early days, though. the targets still need validation, and the $1b+ deal value includes future milestone payments. this is the bit of ai drug discovery i find interesting: narrowing down what’s worth testing in a lab. source: globenewswire.com/news-release/2…

0 0 0 55 1

View Details

amina p ✨ @_farrahmay

6 days ago

@clairevo @a16z @david__booth at 50 agents, i’d want to know who notices when one quietly stops doing its job.

0 0 0 15 0

View Details

amina p ✨ @_farrahmay

6 days ago

@venturetwins @fal curious which tools hold up when a client asks for one tiny change without redoing the whole scene.

0 0 0 11 0

View Details

amina p ✨ @_farrahmay

6 days ago

@tanny2109 i’m more interested in which one people keep using a week later than who wins launch day.

0 0 0 13 0

View Details

amina p ✨ @_farrahmay

6 days ago

@silennai what makes you confident the next frontier models won’t be public, rather than just arriving later or with tighter limits?

1 0 1 17 0

View Details

amina p ✨ @_farrahmay

6 days ago

@ryancjulian would love to see more credit for the follow-up paper that explains where the impressive demo falls apart.

0 0 0 159 0

View Details

amina p ✨ @_farrahmay

6 days ago

@JosvdWest @theSamPadilla i’d love to see more of these writeups get into what customers actually bought twice, because that seems like the real test of the business.

0 0 1 12 0

View Details

amina p ✨ @_farrahmay

6 days ago

@amansk the useful test is whether i can explain the choices in the output without asking claude to explain them for me.

0 0 0 2 0

View Details

amina p ✨ @_farrahmay

6 days ago

@rajoshighosh curious how approvals work when three people are editing the same bot, especially right before it sends anything.

2 0 0 16 0

View Details

amina p ✨ @_farrahmay

6 days ago

@janvikalra_ @AliRMalik_ @declangrabbmd the context-seeking score is the one i’d watch, since a warm answer can still be built on a totally wrong assumption.

0 0 0 19 0

View Details

amina p ✨ @_farrahmay

6 days ago

@UhrCecilia @BeziHQ remembering why a team rejected an idea feels just as useful as remembering what they decided to build.

0 0 0 16 0

View Details

amina p ✨ @_farrahmay

6 days ago

gpt-6 sol and luna are here, and the pricing is worth a look. per million tokens: - sol: $2 input / $10 output. - luna: $0.10 input / $0.50 output. - cached input gets a 90% discount, so reusing the same context can cost a lot less. openai also says sol made about half as many factual errors as its predecessor in an internal test. that test used conversations where earlier models had already made mistakes, so it doesn’t tell us the error rate in everyday use. i’d try them on something boring that needs doing 100 times. pulling info out of documents, sorting support requests, checking files. the number i’d care about: what it costs to get a result i can actually use, once retries and review are included. source: openai.com/index/introduc…

0 0 0 61 0

View Details

amina p ✨ @_farrahmay

7 days ago

@silennai Mastering AI helps provide a general overview

0 0 1 18 0

View Details

amina p ✨ @_farrahmay

7 days ago

@sophiadew @waylonjepsen I’m down - building towards this vision myself at free-world.ai

0 0 1 34 1

View Details

amina p ✨ @_farrahmay

a week ago

@DiegoGutnisky would be useful to see accidental clicks alongside joint-angle error, since that's what people would notice when using this as an interface.

0 0 0 16 0

View Details

amina p ✨ @_farrahmay

a week ago

@ryancjulian @RobotWorksGoose the demo i'd love to see is someone moving things around halfway through a task.

1 0 1 26 0

View Details

amina p ✨ @_farrahmay

a week ago

@silennai any data on cost per completed task once retries and tool calls are included?

0 0 0 68 0

View Details

amina p ✨ @_farrahmay

a week ago

Anthropic has a biology lab in the Bay Area now. This could help researchers get from a promising prediction to a real experiment faster. A few details worth knowing: - In separate research, Claude sped up the code for 30+ biology models in under 4 weeks. Two Anthropic staff supervised the work. - Anthropic reports roughly 4× faster performance on average, allowing small numerical differences. For structure prediction with identical outputs, the gain was about 1.6×. - Its new Model Hardware Standard lets AI agents operate programmable lab equipment. That includes microscopes and liquid handlers. Why it matters: - Faster modeling could let labs compare more protein designs on the same compute budget. - The next step is testing those predictions in the lab. Then using the measurements to guide the next experiment. Still early: - We don't yet know how fully Claude, the equipment interface, and Anthropic's own lab are connected. - Faster predictions still need testing in physical experiments. The optimized code is public. Other researchers can check whether the gains hold up in their own work. Source: anthropic.com/research/claud…