Joined October 2025
- Tweets70
- Following134
- Followers26
- Likes103
Fantastic article by @Mfmansoo on some of the recent discussions around the new Astra architecture. OpenAI's stepwise improvement in abstract reasoning and spatial 3D generation is being attributed by some external analysts to recurrent (looped) layers.
TLDR is that this architecture would allow the model to reason over its own internal state before actually committing to a token. The difference is kind of like thinking by talking out loud and describing your chain of thought vs thinking through a problem in your head.
In the former, your reasoning is limited to what information you can write down, whereas in the latter you have the entire landscape of your imagination. It's going to be very interesting to see how Astra and its successors do in new pathfinder/wanderer objectives.
Very solid workshop. As is necessary with these kinds of events that aim to cater broadly, the talk itself was interesting but a little bit too shallow for any serious insights. The q/a session was pretty great though.
@AEllisBloor had a funny comment that stuck with me: "SFT is underrated", and @MarioMartone01 shared some really cool experiences on building clustered eval sets and seeking out divergence.
Open models keep getting better. So do the ways teams are adapting them.
On Aug 31, we’re getting builders from EliseAI and @baseten together to compare different approaches. This workshop will cover:
→ When fine-tuning and other post-training techniques are worthwhile
→ When harness engineering is the better approach
→ How to build accurate benchmarks and reliable ground-truth datasets
→ How to hill-climb on evals without optimizing for the wrong signals
→ The performance, infrastructure and reliability challenges that emerge at scale
This essay by @Irregular on the future of ai security is really interesting. First thoughts go straight to how parallel this runs with many modern day international security concerns: how to defend against an enemy that you know will outpace you. Strategies like asymmetric warfare and Fabian tactics can offer very strong matchups. Building out security measures that anticipate breaches and optimize for practical data security, rather than building an unbreakable system.
I’ve long held the belief that the individual unit value of “data” will become far less valuable in the coming future. With every single one of today’s systems, across every industry, running on inferences and producing even more data, the question of whether a reference point exists or not has become far less important, and the question asked is how trustworthy is this reference. When “data” shifts from a commodity to a process, real value will come from provable and traceable integrity rather than the data itself.
An essay from @Irregular on where we are, and where we're going: endstatefallacy.com/
I’ve use the term “data” here pretty loosely as an abstraction for actionable input. The emergence of tokens as a unit makes this even more grounded.
Ty @Alfred_Lin for bringing this article onto my feed
Really bullish on @GoogleDeepMind Gemini robotics. Worth reading up on their vision of embodied reasoning
Will this finally kill context window maxxing?
A senior Anthropic engineer just dropped 12-page PDF on "Graph Engineering" for multi-agentic systems.
The shift: your agents memory dies with their context window. A knowledge graph makes it permanent.
Extract → Resolve → Assemble → Query → Repeat
Every agentic graph has 5 stages:
• Extract: Haiku pulls entities and S-P-O triples. One call per doc. The Pydantic schema is the only training data.
• Resolve: Sonnet clusters "Edwin Aldrin" → "Buzz Aldrin" - zero string overlap - using descriptions as context.
• Assemble: canonical nodes, typed edges, provenance on every triple. One connected graph.
• Query: serialize a subgraph → Sonnet reasons over triples → every answer cites a specific edge.
Plug this into multi-agent systems as shared memory.
Workers write to it, evaluators fact-check against it, loops persist overnight with it.
This 12-page PDF changed how I'm building multi-agent systems today.
Read it now, then explore the article below.
There have been some interesting interactions with the new @Kimi_Moonshot k3 model, namely that if you ask for its name, it will claim to be Claude. And often times if you try and course correct, it will persistently claim to be an anthropic model.
A colleague of mine @deankeinan noted that he saw totally different levels of reasoning, and joked that it feels like one in every few prompts is load balanced to anthropic.
Would be interesting if we're really seing distillation done at this level, my first finger point would actually be some kind of student teacher training technique and a gapped routing network.
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.com/blog/kimi-k3
I have been a fable truther even after the 5.6 drop, but the way anthropic is positioning their releases and usage limits is pushing me to codex as my daily driver.
Neil Sachdeva retweeted
LangSmith for Startups Spotlight: @finchlegal
✅ Powers pre-litigation for plaintiff personal injury firms, pairing experienced PI professionals with purpose-built agents.
✅ 10x growth over the past year, powering pre-litigation for leading plaintiff firms across the country.
✅ Uses LangSmith as their production observability layer — tracing, evaluating, and monitoring quality and cost across workflows like treatment check-ins, client communications, medical-record follow-up, document processing, and more.
Want to help build the future of consumer law? Finch is hiring exceptional engineers and operators in NYC: finchlegal.com/careers
It’s about time they added regular chat to the Codex app, wonder to what level they were losing usage
The codex 5.6 window is open:
I have 2 hours between now and the France Morocco game starting, to finish the past week of tech debt I accumulated after hitting my fable limits
