
Founder @bubblspace — AI agent traffic intelligence. At FBA Lab 🧪 I build, train & run real AI models in real code. 👇 https://teapawt.org/t.co/mQsXyw7WIg
Joined April 2011
- Tweets8K
- Following473
- Followers487
- Likes17K
FireHacker retweeted
Looped models are weird. A layer can write to itself or read from "future" layers.
@thinkymachines gave us early access to Inkling 2 Mini so we could study its inner mechanisms. We uncover how it coordinates information flow across layers, blocks, and loops.
Studying Inkling 2 Mini gave us cause for optimism: as transformers continue to evolve, we believe our ability to understand their internal mechanisms can evolve with them.
Read the full post: goodfire.com/research/interp…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
FireHacker retweeted
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
FireHacker retweeted
New Inkling 2 Small to run at home!! 26.2B, 3.4B active
huggingface.co/thinkingmachi…
It's out: Thinking Machines' Inkling 2 Mini weights are public on Hugging Face, ahead of any announcement 👀
• 26.2B total, 3.4B active: a looped MoE whose middle 12 layers run three times
• Text and image in, 1M-token context
• Apache 2.0, one ~20GB checkpoint (NVFP4 + MXFP8) that runs on a 32GB Mac or a DGX Spark
• Its model card puts it at 20.1 on the AA Index, tied with Qwen3.8-27B, and 48.2% on DeepSWE
huggingface.co/thinkingmachi…
FireHacker retweeted
We’re releasing all 57 intermediate checkpoints from Marin “Mochi” 535B-A23B pretraining, including 10 from long-context extension. 18.2T tokens, about 61 TB of weights.
huggingface.co/open-athena/M…
Blog: openathena.ai/blog/marin-moc…
FireHacker retweeted
I want to thanks @SemiAnalysis_ for the motivation.
In the Gemini RL scaling team, we printed & put that article on our desks to be reminded daily of what used to be true at the time.
But the times they are changing :)
FireHacker retweeted
The evals are not ok
New NeurIPS paper: Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking.
Evals drive deployment, safety standards, and research. Yet we ignore noise from prompt, judge, agent harness etc, & so underestimate error in benchmarking & annotation, & it gets worse with more data.
I'll get to what this means for CIs & how to fix, but first let's talk about how this makes benchmark hacking easier 🧵
FireHacker retweeted
DIFERENCIAS ENTRE MORGOTH Y SAURON
¿Era Sauron una simple copia menor de su maestro? En sus escritos tardíos, Tolkien explicó que ambos tenían naturalezas, ambiciones y filosofías opuestas.
Abrimos hilo comparándolos cara a cara según el lore 👇🏻
FireHacker retweeted
Weight Space Programming: a glimpse into continual learning.
Agent edited Qwen 0.8B's weights to inject facts about my X profile directly into the model: my handle, bio, GitHub, and what i build. It edited 2% of the network (5 of 488 tensors). No training, no gradients.
Result:
Before, it had never heard of me ("I don't have access to private information…").
After, it tells you who i am, my bio and GitHub, answer project questions it was never trained on. Everything else it knew is untouched.
Reproduce : Link to reproduction repo in comments. Point your agent to the repo and teach model your own facts.
Andrew Huberman on the best advice he ever got: figure out the most real work you can do every day, consistently, and find a way to reset every week.
"The best advice I ever got: there was a neurologist out at UC Berkeley and UCSF, Bob Knight, some years ago, very talented. I asked him for advice starting graduate school."
"He said, 'Figure out the maximum amount of real work that you can do every day consistently, and then figure out some non-destructive way to reset yourself every single week.'"
"I said, 'How do you reset?' He said he likes fishing. I like to hike. I like to get into nature and spend time with people while I do it. I figured out I could work X number of hours per day, 5 days a week, and I just stuck to it for 25 years."
"People who take a ton of stimulants and just blitz, blitz, blitz, blitz, blitz. These are the same people that crash and disappear. You have to figure out and build your capacity."
"Yes, you have to work extremely hard. You have to be extremely focused. But you also have to figure out how long you can keep this up."
@hubermanlab
Stanford neuroscientist Andrew Huberman built one of the most popular podcasts in the world by turning health research into protocols millions of people actually use.
His top three habits are surprisingly straightforward: get bright light shortly after waking, resistance train, and make the sleep you can get as high-quality as possible.
He joins a16z’s Daisy Wolf to walk through the protocols he actually follows, from his 35-year workout to the lockbox he keeps his social media phone in, and explains why packing in every protocol can backfire if it gets in the way of the rest of your life.
For founders, he explains why the hours your brain does its best work are worth guarding like a fight, how to find the most real work you can sustain, and why sleep is where neurotech will win first.
0:47 Why zero drinks is easier than two
4:39 Meditation means watching the chaos
6:25 Sunlight first, caffeine if training
7:20 The workout he's done for 35 years
10:09 Two sets to failure, low volume
13:11 Find foods that make you feel good
15:55 Why eating too clean backfires
17:29 Why non-sleep deep rest beats a nap
19:05 A sunset habit from a sleep lab
21:18 Who needs blue-light glasses
23:32 Every supplement he takes
30:46 Why his phone lives in a lockbox
32:40 His top three habits for health
35:47 Guard your two daily focus windows
42:47 Neurotech that puts you to sleep
46:38 The best advice he ever got
YouTube: youtu.be/Ezi-nwe6LlY
@hubermanlab @daisydwolf
FireHacker retweeted
Thousands of companies died at the altar of the internet for one business model to survive, the ad. The agentic internet will need a different one.
@kumarharsha2212 and I spoke to The Economic Times about why you can't upsell an agent, vertical Al built for India, and how this wave compares with the early internet.
AI agents that shop and book on a person's behalf change who sees an ad, and who pays for the internet.
Lightspeed Partners @kumarharsha2212 and @MohapatraHemant joined Himanshi Lohchab in @EconomicTimes podcast Corner Office Conversation to discuss which purchases agents will take over and which will stay with people, why an agent cannot be upsold, and why India may prefer to pay for an outcome rather than a subscription.
They also discussed vertical AI built for India beyond its largest cities, how this wave compares with the early internet, and how they use AI assistants themselves.
Watch the full conversation here: youtu.be/KNdl_uhht3g?si=Xzml…
FireHacker retweeted
if you are a foreign investor in india, this is the most comprehensive guide to managing your investment - KYC, demat, transfers, sales, compliances, taxes and more.
@ro_kapoor is brilliant and his firm has been a godsend for us to manage our india investments.
Thoughts from my Shanghai field trip
Sentiment among Greater China investors was far worse than I expected. Unlike the US market, A shares had not seen a single rebound since July. Still holding Chinese and Hong Kong AI stocks that had crashed and Korean memory stocks that had collapsed, investors kept pressing me on when memory would bounce back.
Many investors said the AI trade was over, and a few said they doubted whether even 2027 capex, not just 2028, would go ahead as planned.
Overall, the bearish sentiment was so strong that even I, a bull before coming to China, started to waver.
It got to the point where, faced with the bear case that memory prices could start falling in the third quarter of next year, I took it seriously and couldn't refute it.
As the conversations went on, I gradually began to understand where their pessimism was coming from. In the end, they were all asking the same question.
So what comes after AI coding? What will be the next engine of ARR for Anthropic and OpenAI?
Several candidates came up. AI4S, AI chip design, and so on... But none of them clearly convinced investors.
I couldn't offer a clear answer either. Sure, there are rumors that Anthropic has made major breakthroughs in both AI4S and AI chip design, but what we actually want to know is when any of that turns into money.
With AI4S, there could be beneficiaries right away, such as companies that sell lab monkeys or diagnostic kits. In fact, there already are. But while we can trade those names in the short term, what we ultimately want to know is whether Anthropic can make money by developing drugs. That is still several years away.
As for AI chip design, the TAM itself is hard to calculate, and even if you do calculate it, it is hard to monetize. That makes it difficult for the market to believe in.
Honestly, this feels like the hardest market in history. If a company isn't making money, selling makes sense. But right now the market sells even when companies are making money, even when they beat, and even when they say demand will stay solid all the way through 2028 and 2029.
So my recommendation to investors was to buy memory, and within memory, DRAM. Not because I have conviction in the DRAM cycle, but because I see it as a relatively defensive choice within the AI trade.
Memory stocks are now trading close to 1x P/B on 2027 BPS. Historically, memory stocks have staged technical rebounds when they fell below 1x P/B.
And even if the broader market falls further from here, the three DRAM makers will keep buying back their own shares. I thought that would guarantee a certain level of buying and concluded the stocks would have downside support.
Of course, I don't know whether memory stocks can break above their previous highs. That depends on whether sentiment across the broader market recovers. But I think there is a fairly good chance that memory outperforms the rest of the AI trade.
FireHacker retweeted
🚨Agrani Labs in talks to raise $100 million from Groww founders' Stargazer fund, Google, Peak XV, others
A $100 million investment will mark one of the largest rounds in India's deeptech ecosystem after the likes of Pixxel, Skyroot, Sarvam among a handful others.
Moneycontrol was the first to report on September 1 that Agrani Labs was in talks to raise $50 million (Rs 475 crore), with existing investor Peak XV Partners looking to double down and new investor 360 One joining the round.
That round has now been upsized to $100 million with Groww's founders too coming in, among others.
Scoop with @Goenka_Tushar1
moneycontrol.com/news/busine…
FireHacker retweeted
This is ragebait, do not fall for ragebait.
Always remember, if Ace didn’t fall for Akainu ragebait then Marineford would have been a success.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Mission Success!
GSLV-F18 has successfully placed NVS-03 in the intended orbit.
Congratulations!
Watch live: youtube.com/live/qIiYr3tRSqE
Mission details:
isro.gov.in/Mission_GSLVF18.…
#GSLVF18 #NVS03 #ISRO #NavIC
FireHacker retweeted
Congratulations to our scientists and engineers at ISRO on the successful launch of GSLV-F18 carrying NVS-03.
This mission marks an important step in strengthening NavIC, our indigenous satellite navigation system. Equipped with enhanced capabilities, NVS-03 will further our pursuit of self-reliance in critical technologies and support navigation services that benefit our people.
The dedication and ingenuity of our space scientists continue to make our nation proud.
@isro
FireHacker retweeted
Can you RL design taste into a model?
I explored training Qwen3.8-27B over some open-source design environments.
The answer: you can, but that taste probably won't be yours or mine. And definitely not a real designer's.
jasperlu.com/blog/rl-design-…

































