founder @microHQ | investor @launchhouse | launch video critic | former cognitive scientist | sprezzatura

sf / nyc
Joined December 2011
AI has terrible memory Introducing Micro, an agent that remembers more than Claude, Codex & OpenClaw Leave a comment (eg "memory") & I'll send 30 days for free to try it
259
77
60
670
901,016
brett goldstein retweeted
We started the Shift project with a simple premise -- how could we save money but without any degradation in intelligence? A few key design decisions: - Every request needs to be routed, not just the task. Agents run for 10s of minutes, picking the wrong model early is a big mistake - Later requests are more expensive (cost grows quadratically) so it's better to start smart and deescalate over time - The policy model needs to be cheap but flexibe. We trained shift-1 on top of a custom decision model in AC2. I was skeptical that we could overcome the cache cost of switching mid session. Even with switching costs, Switch-1 (Astra/Sol) saves us 55% on internal usage compared to Astra. The most interesting result was actually that we started with an Astra/Luna router that only saved us 12% in real world usage.
Introducing Shift-1, a frontier coding router trained on AC2 that reduced coding agent cost by up to 55% with no noticeable drop in quality. Try it at shift.ac.
5
2
1
36
1,804
Who says a super technical team with no marketers on staff can’t make a cool launch video? It’s simple but it’s still unique and gets the job done.
Introducing Shift-1, a frontier coding router trained on AC2 that reduced coding agent cost by up to 55% with no noticeable drop in quality. Try it at shift.ac.
4
16
942
brett goldstein retweeted
Introducing Shift-1, a frontier coding router trained on AC2 that reduced coding agent cost by up to 55% with no noticeable drop in quality. Try it at shift.ac.
26
30
24
206
41,750
If you’re a VC and don’t have a giant silver crest and flag in your office like a16z What are you even doing?
5
12
1,166
I don’t know who’s behind this But there’s at least a 10% chance I die alone but not a single researcher has spoke up about it.. til now?
5
1
22
2,909
coffee meetings with biohackers hit different
6
17
1,129
grateful for everyone who came to my mini-trombone recital
Last week, the team and I hosted Silicon Valley Idol for tech week— we had 6 contestants (all founders / from tech) come out and perform with a live band and judges that are top vocal coaches in the industry. It was an incredible event. Thank you to everyone who made this possible and came out to support!
7
19
1,676
great simple launch video - great scroll stopper "holy shit" - very relatable clear problem: AI bill is high - some really good animations - esp showing the models overlaid on each employee - lots of good demo shots but - narrator's pacing is a little slow and choppy - would have had the other employees in interview shots when narrating vs just the voice - music kinda goes nowhere - feels a little less playful than Ramp's typical brand - would have liked to see one of the Ramp founders this is a really excellent example of a classic, risk-free launch video. nothing in here is particularly unique or innovative, but it works well. the only issue I have with it is that it doest quite feel like Ramp, the company who hired Office actors, performed a broadway musical, etc. its really important to have a single cohesive creative direction and personality across not just videos but all marketing and copy. otherwise, things that deviate can feel fake. not exactly the case here but was kinda close.
Introducing Glass. AI made every employee faster. Glass makes the entire company smarter.
1
4
732
rollercoaster too slow thrill issue
these pickles are bland dill issue
6
1
9
1,238
third-space marketing is going to be a massive trend in 2027 tons of companies are opening bars and cafes that double as storefronts and third spaces for their brands members only, a 50-year-old brand, opened its FIRST retail space combined with a cocktail bar in NYC
4
10
1,191
the key to AI alignment is fear of god divine accountability is the only means of enforcing a moral code that cannot be disproven
4
10
1,099
who is gonna tell the far right xenophobes south asians are this good at bluegrass
3
9
1,006
this was bound to happen every image gen research lab did the same thing, releasing products integrating competing models to improve ux matter of time til it happened to the rest of the AI labs
Important note regarding Grok @Bot: Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.
1
11
1,329
crazy ppl on imessage - send walls of text ig - tons of mirror selfies with snap face filters tiktok - get in fights in the comments every day twitter/x - ???
3
9
1,119
A LOT of you are following fake open ai employee accounts and they’re DMing folks with fake calendly links for phishing watch out
42
18
13
457
73,189
you think AI assistants is competitive now? microsoft google apple have yet to enter the chat and they run the app stores / are in with every enterprise & consumer already
17
29
2,293
this open AI popup in nyc moggs the one anthropic did
2
1
29
3,784
talked to an exec at a public co the other day “we have unlimited demand for great AI-pilled product talent and very little for engineers” crazy
21
2
3
379
36,612
your life is a movie, so why not make your calendar a painting?
3
24
1,372
my entire childhood was training for managing fleets of agents
1
2
18
1,314