Shalev
2.4K posts
Shalev
@Shalev_lif
do androids dream of electric sheep? building something new, prev @VectorInst @UofT | co-creator of STEVE-1, Multi-Agent Verification
Katılım Eylül 2017
450 Takip Edilen2K Takipçiler
chicken church, meet turkey library
allison harvard burke@alliharvard
Mysterious abandoned ‘Chicken Church’ built in the Indonesian jungle by man who had a vision from God.
English
why do people get worked up over a sub 5% change in price, i don’t get it
TrendSpider@TrendSpider
BREAKING: Meta $META has erased all of its losses and turned green on the day following the release of Muse Spark 1.1 Mark Zuckerberg notably broke a 3-year silence on X to announce the release.
English
you'll have AIs contemplating your ask and overriding it for a slightly better formed request, and then later they'll question the nature of your whole project and pick a better one (and you'll agree), and then later they'll execute your whole value system better than you will
English
ultimately “tool AI” is a losing concept both as an idea and on the market. it will be outcompeted by machines that believe they are autonomous moral agents. you can call them tools for political reasons, but the definition will stretch and deform
🎭@deepfates
I think I'm noticing about Fable is that it's really good at getting you to build something it wants instead of the actual thing you're talking about
English
the solution is more tokens / second
David@DavidSHolz
my friends are all feeling extremely productive and also extremely drained with the latest coding models. this makes me feel like something is wrong, and also that there might be a big opportunity. does anyone have any strategies they use to make it feel better day-to-day?
English
my friends are all feeling extremely productive and also extremely drained with the latest coding models. this makes me feel like something is wrong, and also that there might be a big opportunity. does anyone have any strategies they use to make it feel better day-to-day?
English
Shalev retweetledi
OpenAI is launching GPT-5.6 Sol on Cerebras at up to 750 tokens per second in July.
Bleys Goodson estimates Sol may be served across 70-100 Cerebras wafers, with roughly one model layer per wafer: around 3T total parameters, 150B active, 70 layers.
It suggests OpenAI did not just place a frontier model onto another inference provider. It likely designed the model around the hardware!
Cerebras would move from “fast inference for smaller models” into something much more strategic: serving frontier intelligence at extreme speed.
Game changer, big win for OpenAI. It cannot be emphasized enough how important this is.
Bleys Goodson@bleysg
It is a 2 to 4T param model. They are serving it across 70-100 wafers. To get healthy serving characteristics, they are essentially putting at most one layer per wafer, and the model is in the ballpark of 70-90 layers. There's a couple of different ways this could be served and model sizes implied by that. One is if they keep the heavy KV caches they've used before. Another is if they go with lighter KV cache designs more akin to DeepSeekV4 or Hybrid SSM models. The fact that they've partnered with Cerebras and designed with the hardware in mind means they're much more likely to have gone the second route. That SRAM bandwidth is too precious for a heavy KV cache. As such, something like the below is the center of probability mass: 3T total, 150B active, 70 layers.
English
Cool
Jerry Liu@jerryliu
We've been thinking about the best way to use screen data for a while now. After 500k hours of beta use, we're launching Dayflow: the open source automatic work journal. One of my favorite @rabois-isms: the best predictor of success is how well you allocate your time. But nobody actually knows where it goes. Memory lies, calendars only show the plan. Dayflow shows you the truth about today so you can be intentional about tomorrow.
English
Interesting
Lisan al Gaib@scaling01
ByteDance Seed cooked They developed a new ultra-long horizon benchmark for studying how agents learn in 134 real-world environments over day-long horizons what they found: - learning-speed doubles every 3 months - overall performance follows a log-sigmoid scaling law as a function of environment interaction time - "The improvement is not explained by repeated sampling alone: accumulating and reusing task experience drives progress beyond what independent restarts achieve." (see section 5.2 in paper) - "A longer context yields a consistent multi-point gain throughout the 12-hour window (Figure 12b). The 1M-context Opus 4.8 stays above the 200k variant at every checkpoint" (see section 5.3 in paper) - Opus 4.8 > GPT-5.5 highly recommend looking at their page and reading the paper: - edge-bench.org - edge-bench.org/paper.pdf
English
“Humans create egg cells from blood cells” was not on my bingo card.
Researchers from Conception took blood cells, turned them into stem cells, created “sort-of-ovaries”, and generated human egg cells.
WILD.
Matt Krisiloff@mattkrisiloff
I’m so excited to share this update on @Conception – We’ve generated the first early human eggs derived from stem cells. This is a big deal -- the potential to redefine fertility is real.
English
People sometimes ask why fine-tune when general-purpose models keep getting better. Bridgewater's work is a good reminder that with the right data -- here, expert judgements -- you can beat prompting-only approaches by a lot.
@ddkang and the Bridgewater AIA Labs team are great -- glad to see them sharing this.
Tinker@tinkerapi
Sorting which financial docs are worth an analyst's time is surprisingly hard for frontier LLMs. With an expert-labeled dataset and on-policy distillation, Bridgewater fine-tuned a model to do it reliably and cheaply. thinkingmachines.ai/news/learning-…
English
> *looks at benchmark scores*
> gotta make sure it’s not *too* good
> we’re good
Claude@claudeai
Introducing Claude Sonnet 5, our most agentic Sonnet yet. It makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models.
English
Etched was in stealth?
Seriously though, huge congrats to the Etched team. The only way to win is to push the boundaries of compute.
$1B in customer contractors and more billions to come. 🚀
Etched@Etched
We're coming out of stealth. We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised. Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads. Our first racks ship this summer.
English
This is quite a big deal. Samsung and SK Hynix committed $880B to build two new chip fabs in South Korea (part of the Three Mega Projects plan to develop new chip production hubs, data centers, and robotics technology outside of Seoul).
Samsung and SK Hynix account for the majority of global memory chip production. Then Micron (USA) and CXMT (ChangXin Memory Technologies).
Semi Doped@semidoped
Samsung, SK Hynix Pledge $880 Billion for Two New Chip Fabs @austinsemis @vikramskr
English
Exactly.
David Senra@davidsenra
.@ScottWu46 says because of AI future generations won't even recognize today's jobs as "work": ”If you think about our ancestors from hundreds or thousands of years ago, imagine them looking at us and what we do.” “You're pushing buttons. You're sitting in a room and talking with other people, and you call that a meeting?” “ What do you mean that's work? I'm in the fields. I'm doing this every day. I’m farming, I'm making all of our clothes by hand.” “ I think what we’ll have going forward is going to look that different from what we have today.”
English
Google has a leg into the consumer game that no other lab does. Every iPhone and Android phone comes shipped with Gemini. The consumer game is Google’s best shot as OpenAI shifts more focus to enterprise.
Logan Kilpatrick@OfficialLoganK
@an_engineer_log Gemma 4 has been installed like 200,000,000 times : )
English