/sesh/null

656 posts

/sesh/null banner
/sesh/null

/sesh/null

@nerdsane

VP, data-and-ai@datadoghq | peripatetic | minimalist { engineer | athlete | artist } | I opinions-my-own

New York, NY Katılım Aralık 2017
816 Takip Edilen262 Takipçiler
/sesh/null
/sesh/null@nerdsane·
I keep coming back to this. Maybe an agent first version of the moby project @solomonstre as an operational model for software. Share components as primitives and assemblies as factories. Some examples that have a solid foundation towards this model similar to ghostty: 1) DataFusion as LLVM for query engines (@andrewlamb1111 ) 2) recently turso (@glcst ) as LLVM for databases 3) Vortex as the LLVM of storage formats (@willmanning )
Mitchell Hashimoto@mitchellh

x.com/i/article/2041…

English
1
1
8
839
/sesh/null retweetledi
arni
arni@arni0x9053·
Kimi K3 @Kimi_Moonshot is a big leap from Kimi K2.7 Code visually. However, it so far has shown to be as slow, to the point that it would be a deterrent for me. And it’s not just tok/s but the amount of turns and reasoning it did, which means it’s also isn’t particularly cheap. I used Grok Build as a harness for this test, inclined to re-test with OpenCode or Pi in case it does better in those.
arni@arni0x9053

Gpt 5.6 Sol for UI expands on gpt 5.5's preference for chunky designs and bold fonts. It especially now seems to prefer alternating background colors to section landing pages, and is overall heavily inspired by graphic design tradition (asymmetrical grids, bold typography, color blocking). I'd say this is a rather unique feature gpt has compared to other models.

English
3
1
4
2.2K
/sesh/null retweetledi
Glauber Costa
Glauber Costa@glcst·
I am excited to announce that we are officially writing a new version of Postgres. In Rust - and creating the LLVM of databases in the process. In the span of a year, we have rewritten SQLite. Keeping the compatibility, increasing its feature set. MVCC, Types, (Live) Materialized Views, among other things. In the process of doing that, we have realized: At the end of the day, what makes SQLite special is that it compiles SQL to a database-specific bytecode. So why can't we compile *Postgres* to the same bytecode? Turns out we can. I ran an experiment called pgmicro as a way to prove this hypothesis, and it works very well. It is time to make this official, and put the weight of Turso behind it. We shall give the world a modern take on Postgres. Wire compatible, but built on a new architecture. We have already heard of others wanting to extend this. MySQL? Redis? the sky is the limit. What can we do if we do for databases what LLVM did for compilers? To prove how powerful the SQLite bytecode is, we are actually running DOOM compiled to the unmodified SQLite instruction set. And because Turso runs natively in the browser, you can play the game in your browser. With the database executing it. Read the full story below! 👇
English
180
384
4.4K
505.3K
/sesh/null retweetledi
Diptanu Choudhury
Diptanu Choudhury@diptanu·
Tensorlake Sandboxes come with a highly optimized storage engine that delivers SSD-class IOPS: 66,100 random-read and 65,000 random-write IOPS at 4 KiB inside the guest, with 4.7 GiB/s sequential writes. Postgres runs within 1% of it on fsync-bound TPC-B. Sandboxes consistently snapshot in tens to hundreds of milliseconds, regardless of how much data the VM has on disk. We wrote a blog post on the design of our custom storage engine in Firecracker, with comprehensive benchmarks of running Postgres in the sandbox.
Diptanu Choudhury tweet media
English
5
11
75
13.2K
/sesh/null retweetledi
/sesh/null
/sesh/null@nerdsane·
Sector regulators like FINRA (finance) or FAA (aviation) are a good mental model for frontier AI oversight. In aviation, FAA/EASA lean on RTCA/DO‑178C to enforce graded, process and evidence based assurance for airborne software. Hopefully we would see a similar, DO‑178C‑style mandatory assurance framework for training and releasing frontier models, replacing today’s largely artisanal practice.
English
1
1
3
580
/sesh/null retweetledi
Bruno Volpato
Bruno Volpato@brunovolpato·
@elonmusk @Alezander907 Yeah, definitely this if possible! Most of our long Claude/Codex sessions are hitting ~95% cached input, so getting 75% discount instead of the 90% OpenAI/Anthropic are offering effectively double the input cost (1.98x).
English
0
1
5
130
/sesh/null
/sesh/null@nerdsane·
@wmf First I wanted it to happen. Now I’m seeing more evidence of this happening.
English
0
0
2
17
Wes Felter
Wes Felter@wmf·
@nerdsane You think that is happening or that's what you want to happen?
English
1
0
0
60
/sesh/null
/sesh/null@nerdsane·
I agree. we can perhaps draw an analogy with photography and drawing. Cameras automated capture and photography became about framing, composition and post processing, but drawing still endured because it trains perception through the hand. Some of the best photographers I know are also very good at drawing. AI may automate software production, but programming small systems from first principles by hand will remain how we train the mind that must understand and judge them.
antirez@antirez

There is another idea that continuously return in my head recently: we should not stop writing code by hand. But this code will be like poetry. Small programs that are outstanding either because of style or new ideas introduced. Code you may print and frame on your wall.

English
1
1
4
515
/sesh/null
/sesh/null@nerdsane·
Whats the best self-serve option to have my own inference endpoint for post-trained qwen 3.6 35B with my writing style and voice using the GRPO approach @ivan_bezdomny and @arni0x9053 are talking about + also RL continuously as I use it as my editor.
arni@arni0x9053

A great write up. I’ve been sensing that Qwen 3.6 35B has a lot of potential for writing (and I actually think front end design too). This result is super encouraging and bring GRPO the top of the next steps in my writing styles/VOICE.md explorations. x.com/arni0x9053/sta…

English
4
0
7
2.6K
/sesh/null retweetledi
Diptanu Choudhury
Diptanu Choudhury@diptanu·
We built a new dataplane L4 proxy for @tensorlake sandboxes, using kernel TLS offload and splice which gives every connection 2.5 GB/s into a sandbox from external networks. This improves user perceived performance of agents which create or ingest large files. tensorlake.ai/blog/near-zero…
English
0
5
29
1.9K
/sesh/null
/sesh/null@nerdsane·
VOICE.md (like DESIGN.md but for writing) more and more.
arni@arni0x9053

In the age when everything sounds like Claude, I have been looking into the concept of VOICE.md (like DESIGN.md but for writing) more and more. I majored in philology and linguistics "back in the day" (Mandarin and English). I operate most of the time in a language (English) I learned later in adolescence which means I inherently rely especially heavily on pattern matching in it - a late-acquired language is less deeply entrenched than a childhood one, and thus more susceptible to whatever dominates recent input. I also work solo most of the time which means the ratio of me talking to Claude/gpt is high, and that is a dangerous combo. So paying attention to style and writing is at the top of my mind. Some might even say it is "load-bearing" (jk jk). How do you find your voice and don't lose it? How do you extract and define it? What should be included in the profile and VOICE.md? How do you get an LLM to actually produce prose in your style? We've been exploring the style adherence part with Fable as groundwork for writing styles in katagami.ai (which I will be launching soon alongside design languages), so far we're landing on this: Take 10+ samples of human writing (corpus) -> extract its measurable habits: sentence rhythm and its spread, punctuation rates, function-word profile, character-level texture, vocabulary richness -> produce VOICE.md (containing never-rules, five verbatim gold samples, signature vocabulary, the measured fingerprint, and the machine contract - every numeric limit computed from the corpus) -> ask an LLM to output text in this style -> run deterministic stylometry checks on the output (their thresholds come from the author's own out-of-sample variation, so a text only fails by deviating from the author more than the author deviates from themselves) + an LLM coach giving directional adherence feedback -> LLM-writer produces corrected output that is closer to the author's own measured habits (sentence rhythm, word frequencies, punctuation). Measured by: - a word-frequency fingerprint - an openly released authorship-embedding model (STAR) - human (my) judgement (the first two are validated to tell authors apart; whether their distance scores mean a text reads more like the voice is what I'm still validating) Here's some observations made so far:

English
0
0
1
195