NYC
- Tweets 4,361
- Following 1,188
- Followers 767
- Likes 4,862
300m params is TINY, super impressive how small and powerful it is. one click deploy in your retrieval stack via @mixpeek
mixpeek.com/model/google/embโฆ
Introducing EmbeddingGemma 2! ๐
Our lightweight, multimodal embedding model maps text, code, images, video, and audio into a single, unified embedding space. Optimized for on-device use cases, it features:
- 740M parameter form factor with modular encoders
- Flexible dimension sizes (768dim-128dim) via Matryoshka Representation Learning (MRL)
- 8K context window (4x larger than text-only EmbeddingGemma)
- A commercially permissive Apache 2.0 license
4
129
9/11 was terrible. maybe the US shouldn't have stationed troops in saudi arabia?
the boston marathon bombing was horrific. maybe the US shouldn't have invaded afghanistan?
charlie hebdo was painful. maybe they shouldn't have drawn the prophet?
this is "justification laundering". you concede the tragedy, then make an excuse.
blows my mind that anyone still thinks he cares about anything that isnโt his people or dogma.
This day is profoundly painful for so many New Yorkers. It marks three years since the horrific war crimes of October 7. On that day, Hamas killed more than 1,100 Israelis and abducted 251 others. We grieve alongside our many friends and neighbors who still carry that unthinkable loss.
The suffering did not begin nor end that day. It instead continued through the Israeli governmentโs devastating, ongoing genocide in Gaza, in which it has killed more than 74,000 Palestiniansโincluding more than 21,000 children. Even since the so-called ceasefire, more than 1,400 Palestinians have been killed. That suffering has intensified with every Israeli bomb dropped on ambulances waiting outside hospitals, every drone strike targeting a residential building or journalist, and every shipment of food aid turned away at a border crossing. Many New Yorkers are reckoning with this unbearable grief, and we mourn alongside them โ all while knowing that our tax dollars fund these war crimes.
Three years later, we must refuse to accept a world in which occupation and apartheid continue, where war crimes are met with impunity, and our federal government continues its complicity with every arms shipment.
To honor the dead, we must fight for a world in which every person can live with freedom, safety, and dignity.
6
263
iโm constantly saving instagram posts about events and then adding them to my calendar later, such an obvious AI use case.
so now in the amux iOS app, you can share to amux workers directly in the native iOS share flow.
here iโm sending an IG post about an event to my โsocial-activitiesโ worker thatโs stays alive, has schedulers and its own directive.
it just magically appears.
127
muse is the only browser-based llm that is able to access my tailscale and can send commands within my vpn.
๐จ side project alert ๐จ
Announcing Muse Gadgets, an open source ESP32 firmware and Linux sdk so that you can make hardware devices that work with Muse.
Grab an API token from gadgets.muse.ai and point your favorite coding agent at the github repo to build your own peripherals for Muse.
1
103
building a server from scratch made me realize how bizarre the hardware supply chain is.
i can buy CPUs, ECC RAM, 40TB of storage, a chassis, PSU, HBA, UPS, etc. basically on demand.
But there is no compatible motherboard in sight.
we talk a lot about compute scarcity in terms of GPUs, but boring infrastructure components can be an even dumber bottleneck.
1
75
no offense to anyone in this thread, but it perfectly captures what happens when you hop on a technology trend without understanding fundamentals.
of course an LLM can regex + grep and continue scrolling until it finds what it's looking for.
is this efficient? absolutely not, this is why databases, indexing technology and search exists to make it more efficient!
1
2
221
public safety footage has two search problems wearing one trench coat: "find the moment" and "prove who was allowed to see it". multimodal evidence only works when retrieval knows who may see what, down to the second inside a file
47
a brand protection team found 400 counterfeit listings in 3 minutes. the detection model was the easy part. the hard part was indexing the whole marketplace so "looks like our product" is a query you can just run
2
109
it may seem like every company: Meta, Anthropic, OpenAI are all vying for the same GTM talent pool, and it's because they are.
MongoDB vs Meta and Factory vs Cognition... it's all fighting over GTM leaders who drank the "playbook mafia" koolaid.
ok so Snowflake, Grafana, MongoDB, Zscaler, Rubrik, and Datadog, what do they have in common? they all teach their pre-sales teams how to sell in a formulaic way.
It makes sense when you think about it because they all sell to engineers, and engineers love to measure things.
this "playbook mafia", is nothing more than cohorts of AEs, SAs, CSMs and PS that have drank the koolaid of systems-based selling.
there are plenty of flavors of this koolaid: Command of the Message, MEDDPIC, Challenger, etc. but they all have the same characteristics which is why these sellers are so valuable.
109
RAG has gone thru ebs and flows:
v1: tokenizers splitting by word, phrase etc (tf-idf)
v2: chunk a corpus by sentence then embed each sentence (knn)
v3: query intent and reranking (llm)
v4: hierarchal agentic search (agents)
v5: learned indexes and learned rerankers
weโre currently in between v3 and v4, all signs point to v5 and the research shows it
2
4
419
reciprocal rank fusion is the least glamorous thing in our stack and it decides every result. retrievers disagree about scores, so rrf ignores scores and blends by rank: 1/(60+rank), summed per document. you can mix bm25 with clip embeddings and never normalize anything
60
a vcโs real job should be finding people already pointed at their lifeโs work.
failure is not an option for them, it would make the last decade look like a waste.
they also operate like cockroaches: low burn, high persistence, default-alive.
that combination is rarer than the purest missionaries or mercenaries, and it survives talent-poaching.
Two random honest thoughts on the market right now..
First is on people: It's harder than ever to find missionaries. I'm seeing people leave the companies they have cofounded to join the "obvious" winners. Public CEOs are leaving for positions at labs. Founders leaving and rejoining neolabs at a moment's notice. Companies getting Windsurf-ed where only the choice people are getting to join the future company. Founders raising tranched valuations without worrying about the next employee that will join them will have 409As that make early exercising impossible.
As early stage VCs, we are often just betting on people. And, when those people leave on a dime's notice, then the remaining company (usually) becomes a shell. It seems like a great time to be a mercenary, and everyone feels like one even though there are missionaries out there.
Second is on defensibility: There's a lot of chatter on moats, so I won't re-litigate the matter here. But the thing that I want to call out is given the pace of what's happening the moat is no longer "one" thing. Like if you really think about it especially in the app layer - the harness is like a thousand small things done well. And, a lot of it is invisible. And given the competition is so fierce and new entrants overnight, you really have to dig into the details to figure out why a given company is different. What is their edge, and their combined edge across product, tech and GTM.
And often times even with that, you end up going back to "well the real defensibility is the founders". And in those founders the defensibility is their speed of execution and being able to adapt to the market. That has become the new bar for every founder out there, given there are no "blue ocean" markets. And even if there are, any company with a LITTLE bit of heat, will have a lot of competitiors overnight.
Even though I think I have a pretty good sense on finding the best people who have founder market fit, I find this current climate really hard and it's testing all my priors (and patience) at the same time.
I know it's not true, but feels like everyone is in it to grab the bag and not really care about their legacy of what they want to build (please reach out if you do).
2
2
239
lord give me the courage of the chabad jews in union square that ask every jewish looking person if theyโre jewish. iykyk.
luckily i give them a hint.
5
2
194
3,671
perfect example of how ai should be deeply integrated into existing ux patterns.
i despise the chat bot interface, more adoption of this please.
2
3
718
when I joined mongodb in 2018 as an enterprise solutions architect, I was looking to learn sales from the best.
i had built a pile of side projects as an engineer. I could ship. I could not sell. mongo was the first place that made the enterprise motion feel like a real system.
thatโs why the CJ thing is more interesting than โCEO left after 10 months.โ
mongo does not casually hire a CEO. they were specifically looking for someone who could run an enterprise motion while atlas and atlas search became the ai growth engine.
meta likely waited for mongo to finish the diligence, then hired the person mongo had already certified as โcan sell serious software into serious accounts.โ
mongodb.com/company/newsroomโฆ
1
279
i run my linkedin like a deploy pipeline now. draft in git, grade against a rubric, human QA before anything ships, measure on mondays. the grader rejects below 35/50 and it rejects a lot of my drafts. posting less, landing more
1
66
its exciting that folks are discovering new methods of information retrieval, but this has a fundamental flaw: it doesn't allow you to do "data exploration".
once you use jev (or any NER model) to classify text, then store that text... you're then bound by that representation that you've stored.
saying "RAG is cooked", is obviously clickbait because not only is this still RAG, but traditional R in RAG is approximate nearest neighbors.
said differently, it allows you to search for something you haven't stored classification for verbatim.
now, the answer to this is to encode the classification from jev but then we're right back where we started.
i am however bullish on using classifiers like jev for agentic search however, and I'm happy that @mixpeek was one of the first to adapt it for our customers.
Been wanting people to use Jev for this for a while now :)
Semantic hierarchical search! just like a human would. check out our hierarchical classification cookbook
1
1
1,072