Alex Pareto

Building.

Previously Head of Spend Management Engineering @brexhq, founder Demeanor (@ycombinator S18), NTWRK, @facebook

[Alex Pareto @alexpareto]

@jamesreggio Must be 80%+

Though I think the review will still be there -> but shift to more of a product review

[Alex Pareto @alexpareto]

Today's PR review interface is far too flat... and it totally breaks with how much code we are creating now. We need an interface that differentiates between core business logic (requires human review!) from dependency plumbing, test mock setup, and other skimmable changes.

[Alex Pareto @alexpareto]

The more interesting part about this I think is the far right - almost every model is at the same performance tier with a good Agents.md. With good prompting and context a lot of these models are starting to feel incredibly similar.

[Guillermo Rauch @rauchg]

Kimi K3 is the best performing model on https://t.co/aporqgIfIh, ahead of Fable, reaching a comparable success rate in less time.

This is the first time that an open model is ahead of all proprietary ones for this comprehensive web engineering benchmark.

Notes:

▪️ Benchmarks don’t always tell the full story, although this is important signal, adding to mounting evidence that this could be a breakthrough moment for open models

▪️ No model as of yet has reached 100% completion on this set of evals. The top performer peaks at 92% and 96% “with help”

[Alex Pareto @alexpareto]

@jackbutcher Yes

[Alex Pareto @alexpareto]

@alialobai1 Moving our codex runs to a cloud VM was a huge unlock honestly (ours was getting bogged down w parallel test runs)

[Alex Pareto @alexpareto]

One of the biggest unlocks for our engineering velocity has been moving all coding agents to running in the cloud - everyone gets their own VM to run agents on.

Tests run faster, agents work overnight, and there's no bottleneck on parallel agents.

[Alex Pareto @alexpareto]

We've done the same over time - our integration strategy has quickly gone from:
1. iPaaS
2. SDK
3. Raw API

The time savings from an SDK have rapidly been eclipsed by the debuggability and flexibility of the direct API.

[Alvin Sng @alvinsng]

https://t.co/AUKp96JRpQ

[Nathan @thisisn8cha]

Thinking Machines just released their open weight model, trained on GB300s with an NVFP4 checkpoint on day 0: https://t.co/fBvpcge4qp

NVIDIA themselves are probably the most prolific American provider of open source models (878 models and counting), all of which are optimized for the CUDA stack. Fast training cycles on the best chips, optimized for cost-effective inference, ready for enterprise customization and control.

America is not only running the frontier capability race.
America is also leveraging its hardware strengths to compete in open source!

[Alex Pareto @alexpareto]

Codex's context compaction is quite impressive. Rarely do I ever see a compaction event degrading performance significantly. This started with GPT 5.5 and has only become stronger with 5.6.

Huge step up from when compaction basically meant your thread was over

[Alex Pareto @alexpareto]

@RaminNasibov one of my dreams is a plane with bunk beds that only flies 8 hour red eyes

[Alex Pareto @alexpareto]

@Vtrivedy10 Yes, the taste in what is worth eval'ing, what is worth adding to prompts, and how to phrase those prompts is far superior compared to 5.5 and Opus

[Alex Pareto @alexpareto]

Early reaction - but Sol is pretty incredible at writing and improving evals. Opus 4.8 and 5.5 both have a tendency to "eval hack" by writing evals where the solution is in the prompt itself. Sol feels like the first model that thinks in a software 3.0 mindset.

[Alex Pareto @alexpareto]

There's lots of subtle differences between "agent native" and traditional b2b software.

Example: traditionally the entire RBAC/audit log data model has no concept of an agent acting on behalf of a user. Whereas in an agent native system this is the vast majority of actions.

[Alex Pareto @alexpareto]

Paying for hosting/reliability/enterprise support on the other hand is a whole different ball game. Could imagine someone doing this for OpenCode

[Alex Pareto @alexpareto]

I wonder if we’ll start seeing a shift to open source coding agents just like we have with open source models, databases, and code editors. Paying for a closed source harness/agent reminds me of paying for Oracle DB. Postgres ultimately won, but it took time.

[Nathan @thisisn8cha]

The public narrative around American vs. Chinese models is misleading.

Winning on benchmarks and X/LinkedIn mindshare doesn’t last very long, and more importantly, it doesn’t resonate with those outside the realm of tech elitism.

Having spent the past 6 years closing gaps between AI research and product deployments across verticals like manufacturing and cybersecurity, I firmly believe the real battlefield is product design and change management.

Researchers can and should continue pushing the frontier, but the rest of us should shift our attention towards designing products that the whole world loves. That’s where the economic victory is.

[Alex Pareto @alexpareto]

One of the best parts of the latest OpenAI release is the naming. Sol, Terra, Luna is so much easier to talk about than Pro, Normal?, Mini, Nano.

This was an area Anthropic got right with Opus, Sonnet, Haiku - nice to see OAI follow suit.

[Alex Pareto @alexpareto]

ChatGPT's new voice mode architecture looks very similar to Thinking Machine's launch back in May. A small, fast voice model in front to keep the conversation fluid and natural - and a workhorse model in the back to actually do deeper work.

So many voice agents out there today are backed by IFTTT graphs because low latency models can't handle complex agentic workflows. By pairing a frontier model with a voice model sitting in front - you get the power of the frontier model but the feel of talking to a person.