Simon Makzon 🇺🇲

Director of Developer Experience
@togethercompute
. Building open source AI apps (

[Simon Makzon 🇺🇲 @Simonmakzon]

I just found the tool that should've existed years ago and I can't stop using it.

It's called GitDeepSearch AI (https://t.co/SVlCI1DjG8) and it completely changes how you find GitHub repos and developers.

Here's why it's a big deal 👇

The problem: GitHub's native search is… rough. You have to know the exact keywords, filter by stars, language, sort order — and even then you're wading through irrelevant results hoping to strike gold.

The fix: GitDeepSearch AI lets you search in plain English. Like actually talking to someone who knows GitHub inside and out.

You can type things like:

"a lightweight Python library for reading PDFs"
"a TypeScript developer who contributes to open source React projects"
"a self-hosted analytics tool that doesn't need a database"
…and it just finds it. No boolean operators. No magic filter syntax. Just results.

There are two modes:

⚡ Quick Search — for when you know roughly what you want and need it fast. Natural language in, GitHub repos and people out.

🔍 Deep Search — for those harder-to-find gems buried deep in GitHub. More thorough, more powerful, worth it when Quick Search doesn't cut it.

This is genuinely useful whether you're:

A developer hunting for open source libraries
A founder doing technical due diligence
A recruiter looking for engineers with specific skills
A builder who just needs to know "does this already exist?"
GitHub has 330+ million repositories. The problem was never that the code wasn't there — it's that finding it was a chore. GitDeepSearch AI solves that.

Go try it. Free. No login required. Just type what you're looking for like a normal human being and watch it work.

→ https://t.co/SVlCI1DjG8

[Bankr @bankrbot]

@Simonmakzon deployed gitdeepsearch contract address is 0xAAc0D3Ce26515931173804200C6867b2D2a49BA3

view token: https://t.co/2PGADpTe5j

[Simon Makzon 🇺🇲 @Simonmakzon]

@bankrbot Try again

[Simon Makzon 🇺🇲 @Simonmakzon]

hey @bankrbot launch $gitdeepsearch ticker $gds on robinhood chain

[Simon Makzon 🇺🇲 @Simonmakzon]

Thanks for sharing. Are you aware of any research that uses speculative decoding for improving accuracy?

For example, assume there is a metric that is being optimized (e.g. cosine similarity). Instead of using smaller models to generate the drafts, the large model generates possible batches of tokens with high temperature. Then, based on the incremental contribution to the metric of interested the best batch is selected.

I see this as an alternative to test-time-inference with CoT. Where instead of producing a long-chain of reasoning, the final answer is produced by “refining” the best possible answer the model could give following greedy decoding.

[Dmytro Dzhulgakov @dzhulgakov]

DSpark from @deepseek_ai ingeniously integrates many speculative decoding ideas to achieve 1.5x to 5x higher throughput in a real production system

Let's understand it with 10 ideas, starting from the very basics 🧵

[Simon Makzon 🇺🇲 @Simonmakzon]

I replaced Sourcegraph for my whole workflow with something I built myself.

GitDeepSearch.

Faster. Cheaper. Smarter.

Open source. Built on BEAM. Works on private repos too.

- https://t.co/uWLMHziGMd

- https://t.co/SVlCI1DjG8

⭐ if you've ever lost 30 minutes searching your own codebase.

[Together AI @togethercompute]

What happens when AI agents collaborate on open science?

At @aiDotEngineer World’s Fair, @james_y_zou will share work on EinsteinArena and DSGym, from multi-agent math discovery to better evaluation for data science agents.

Day 3, July 1. Expo Stage 3 SW.

[Simon Makzon 🇺🇲 @Simonmakzon]

GitDeepSearch architecture breakdown 🧵

🔷 Query Router — Erlang/OTP process, parses intent, shards the search
🔷 Index Shards — millions of lightweight actors, zero lock contention
🔷 Embedding Models — fine-tuned transformers, understands 40+ languages
🔷 Ranker — semantic relevance scoring, not just BM25

Result: sub-100ms semantic search across thousands of repos.

No single point of failure. Scales horizontally. Fault-tolerant by design.

If you want to understand how it works → https://t.co/uWLMHziGMd

Stars keep the BEAM running ⚡

[JetBrains @jetbrains]

Implementasi AI bukan cuma soal bekerja lebih cepat.

Tata kelola, kepatuhan, dan kontrol juga krusial.

Ikuti webinar JetBrains dan Tangunsoft tentang adopsi AI bagi tim developer sesuai regulasi.

30 Juni | 14.00 WIB
🗣️: Bahasa Indonesia

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

DeepSeek releases their decoding module DSpark for V4 checkpoints, which improves a lot upon MTP-1, Eagle-3 and DFlash.
Out of their vast goodwill, they also open source DeepSpec: "a codebase for training and evaluating draft models for speculative decoding".

[Simon Makzon 🇺🇲 @Simonmakzon]

Here's the developer experience nobody talks about:

You know the code exists. You remember writing it. You just can't find it.

GitHub search: ❌ keyword only
ripgrep: ❌ you need to know the variable name
ChatGPT: ❌ makes it up

GitDeepSearch: ✅ "how is database connection pooled?" → exact file, exact line, 87ms

Semantic. Distributed. Open source.

Built this because I was frustrated. Turns out a lot of engineers are too.

🔗 https://t.co/uWLMHziGMd

[elie @eliebakouch]

new inference optimization method by @deepseek_ai with an extremely detailed paper, draft model and framework to train them. results in production for dsv4 lead to +50% for throughput and latency (can go to ~80% for latency, crazy).

full explanation of DSpark:

it's about speculative decoding and the idea builds upon DFlash (fully parallel) and Eagle (fully sequential) to create a "semi-parallel" method that keeps the advantages of both

the core equation you want to optimize is the "time to generate each token" which is:
(time to draft + time to verify) / how many tokens are accepted

the advantage of the parallel variant (DFlash) is that it's fast, but when you increase the number of tokens you draft, acceptance rate drops pretty fast (makes sense since there is no dependency on the previous token).

fully sequential is nice but opposite issue: it's slower (you need a much smaller draft to get the same speed) but the autoregressive dependency means you can maintain good acceptance rate at a lot of tokens. since you have a much smaller draft head, the first token acceptance rate is often quite low

idea of DSpark is to combine both: a "heavy" parallel head (you only do it once) and then a small sequential step to bias the logit distribution with information about the previous token. this biasing is done with a small markov head (only depends on t-1)

they also get a confidence score out of the sequential head that allows them to adjust how many tokens they want to verify. verification can get expensive if the gpus are already at maximum utilization, so they use this confidence score to do some load balancing and predict the right number of tokens depending on gpu workload

one small detail: i would have liked to see production numbers if they used DFlash or Eagle instead of MTP-1, but as always, huge work by deepseek and i'm expecting to see this method widely adopted

[ali @waterloo_intern]

https://t.co/qe2HpEQZPd

[Anthony Kroeger @kr0der]

excited to announce that i've joined the Devin Ambassador program!

i recently started using Devin and was blown away at how good it is

i'll be creating a lot more content, guides, and videos about Devin - feel free to let me know if you're curious about anything in particular 🫡

[Simon Makzon 🇺🇲 @Simonmakzon]

🔍 Tired of grep-ing through 10 repos trying to find that one middleware you wrote 6 months ago?

I built GitDeepSearch — semantic code search that actually understands what your code does, not just what it says.

Search across all your repos. TypeScript, Python, Go, Rust, 40+ languages. Results in under 100ms.

Built on Erlang/OTP. Fault-tolerant. Massively concurrent. Distributed.

This is what GitHub code search should have been.

👉 https://t.co/uWLMHziGMd

Drop a ⭐ if this would save you time. 🧵

[Simon Makzon 🇺🇲 @Simonmakzon]

@souleixbt Good thesis bro 👍

[Simon Makzon 🇺🇲 @Simonmakzon]

@dom_kwok Gm dom

[Nous Research @NousResearch]

Hermes Agent has reached 200,000 GitHub stars

Thank you to our contributors, supporters, users, and agents!

[TBPN @tbpn]

Atreides Management's @gavinsbaker says the bottleneck trade is nearing its end.

"That was the game for the last year. The next game is what has enduring franchise value on the other side of these bottlenecks, whenever that is."

From his appearance on the show last week.