Ethan Morrow

AI thinker and tech critic ️ Loves modern art and dystopian sci-fi ️

[Neha Sharma @hellonehha]

it was ,my goto tool for research and for many people in my network. However, I stopped using it way back and never heared about it from anyone.

for me it was becoming overwhelming to use it.

I think they had everything but still somehow I lost the user's interest? or other in compition went ahead of them.

[Niels Rogge @NielsRogge]

Been working on an improved Healthcare & Medical AI page on Papers with Code

The goal is to give people a good overview of the top benchmarks and papers

@iScienceLuvr, would you mind reviewing it and giving some feedback?

Find it here https://t.co/ND7CS6Z8eR

[tensorqt @tensorqt]

tf was up with those deepmind vagueposters

[Shantanu Goel @shantanugoel]

This was meant to be!

No model, and a very superthin layer on top product ripe to be forgotten as soon as Google integrated Gemini into search for free, and then doled out gemini itself for free everywhere. And then chatgpt added their own free plans and even made go free in regions.

All that was left was for them to hit the final nail all by themselves by being scummy and going back on their free offers and charging people without notifying them.

Everything else they launched has been a lot of hyperbole with no substance as well.

Everyone wants to be steve jobs by attitude, but not really be him in execution.

[Matt Dancho (Business Science) @mdancho84]

🚨 Want to learn how to build + ship AI and Data Science projects (that businesses actually want in 2026)?

On September 2nd, I am hosting a free workshop to help you get started with AI + DS projects in Python.

Register here (500 seats): https://t.co/P6jZxC0iBL 

[YARD @notEgoyard]

CLAIMED OPUS 5 COULD ONE SHOT AN ENTIRE MOTION DESIGN VIDEO FROM ONE LINK

one link in, one finished motion design sequence out - no back and forth, no revision passes, first attempt

What actually happened when someone independently reran the exact test:

Same source link, zero extra context, one prompt
Opus 5 produced a fully sequenced motion piece - timing, transitions, visual hierarchy - that held up against professional output

This wasn't Claude Design's UI/UX mode

was raw motion sequencing, the category most reviewers assumed still needed a human directing every beat

Cross reference this against the other honest test running in parallel: uploading a real design.md - Anthropic's own brand tokens - and asking both Opus 5 and Fable 5 to build a 3 screen app from it

Neither model generates its own images inside Claude Design; both still need licensed photos uploaded

That's the one hard limitation surviving across every version so far, motion or static

gap that matters isn't "can it one shot a video" Multiple people already independently confirmed that part

it's what still doesn't scale even at Opus 5 price point

full below

[Kyle Chan @kyleichan]

Major Western companies that have built AI products on Chinese models:
- Thomson Reuters -> Qwen
- Harvey -> Kimi K3
- Cursor -> Kimi K2.5
- Airbnb -> Qwen
- Perplexity -> DeepSeek

[Brace @BraceSproul]

Figuring out how to make an agent "forget" things is a super tricky problem. We just added a new system in OpenWiki to support this. Great read from @colifran_ on how this works:

[InvestmentGuru @InvestmentGuru_]

Nuclear energy
$OKLO
$SMR
$NNE

[Arnav Gupta @championswimmer]

Grok 4.6 only starts off at 150 tok/s for the first few interactions. As soon as context fills beyond 50k tokens it drops down to ~10 tok/s

This is some absolute weird batching setup on their inference end.

[Mike Julian @mikejulian]

Reminds me of the Airtable acquisition, where https://t.co/VQFxflhUY1 was sold but the budding Hyperagent was spun out to remain with the founders

[Mike Julian @mikejulian]

Poolside aside, what’s interesting to me is all the unique outcomes AI has created.

Pre-AI, this would have a simple layoff or a wholesale acquihire.

[YARD @notEgoyard]

OPUS 5 SCORED NUMBER 1 ON ARTIFICIAL ANALYSIS

Humanlayer ran the test nobody at a lab would publish

Opus 5, Opus 4.8, and Sonnet 5, through a 17-checkpoint SlopCodeBench gauntlet

no full task description upfront, requirements revealed one checkpoint at a time, exactly how a real project actually unfolds

scoreboard everyone will screenshot:

Opus 5 is the new #1 on Artificial Analysis, 88.6% on HumanEval/MBPP, a 5.7 point jump over Opus 4.8

number that actually matters, buried under that:

24% strict pass rate across the full 17 checkpoint run - a model that passes checkpoint 1 clean can be sitting in an unmaintainable mess by checkpoint 5 Every prior checkpoint's regression tests still have to pass at every new step - most benchmarks [ SWE bench, Frontier Code ] hand over one complete spec and grade the output once

one grades the same codebase six hours later 5x more degradation showed up under iteration than a single task benchmark would ever reveal

gap between "wins the leaderboard" and "still clean six hours in" is the actual story here

single task benchmark measures whether a model can build something once

Nobody real codebase gets built once

Full below

[Irushi @Im_IrushiK]

Something's seriously wrong with Anthropic’s usage limits.

20x Max + 50% higher limits should basically mean 30x… yet people are burning through a whole week of Fable in barely a day.

30x of WHAT?

$200 for the highest plan and still worrying about usage every day is crazy.

[Sam D'Amico @sdamico]

California prides itself on leading on renewable energy, politicians aspire to grow working-class jobs in the state, and everyone knows that the future involves technologies pioneered here, especially electronics.

Meanwhile, we’ve banned fabricating printed circuit boards here.

[Austen Allred @Austen]

Kelly (my AI agent) is now autonomously building and deploying apps, running tests of a bunch of ads, and scaling out the ones that work.

It’s taken a long time (and a ton of model/harness improvements) but fully autonomous businesses are definitely possible.

[tokenbender @tokenbender]

everything i am reading in a single place - 7

#1 an interesting short read - how to keep thinking.
it made me feel, amongst other things, of why should i write much more for myself and you all.

https://t.co/Z08KO7lcur

[Poe Zhao @poezhao0605]

Zhongji Innolight, China’s largest optical module maker, released its half-year results today. The headline numbers are what you would expect from the main supplier of 800G and 1.6T modules to American hyperscalers. Revenue Rmb 41.8bn ($6.1bn), up 182%. Net profit Rmb 13.65bn ($2.0bn), up 242%. Module gross margin 46.6%.

The cash flow statement tells a different story.

In Q1, Innolight turned roughly 60% of net profit into operating cash. Across the full half-year, that ratio fell to 13%. Operating cash flow dropped 44% to Rmb 1.8bn ($265m). The company attributes it to higher cash payments for purchases. Investing outflows grew more than 8x to Rmb 6.8bn ($1.0bn), driven by long-term asset purchases and external investments. Cash and equivalents ended June at Rmb 6.36bn ($936m), down 10.7% from a year earlier.

The implied Q2 operating cash flow is negative, in a quarter with about $1.2bn in net profit.

I think this is less about Innolight and more about what the AI infrastructure buildout demands from its key suppliers. Record demand forces record forward commitments: prepayments to suppliers, new capacity, receivables from customers who are themselves spending at historic rates. Profit books now. Cash arrives later, if the cycle holds.

Most English coverage of Innolight focuses on profit growth or the Pentagon listing. The cash flow structure gets little attention. I think it is the more useful signal for where this company sits in the cycle.

So here is my question. At what point does a company with $2bn in half-year profit and shrinking cash reserves stop being a growth story and start being a warning signal? Or is this simply the price of being essential right now?

[Philipp Schmid @_philschmid]

Humans prefer talking to Gemini🐸 Gemini 3.1 Flash Live is #1 in the new @ArtificialAnlys Speech Agent Arena. https://t.co/MJumMWaoYi

[Peter Dedene @dedene]

He came from Google to join OpenAI, left to start Thinking Machines, came back in January, and is already out again.

What the hell did Luke see?