@trillhause_ gotta use the snobbery angle
most matcha lattes are made with cheap matcha that’s mixed with powdered green tea, not the high grade stuff cultivated in Kyoto
I’m looking for people in Toronto who can help me pull off the 2nd season of Toronto School of Foundation Modelling. Have a few exciting guest speakers locked in already. I have more ambitious goals for this season. I’m looking for someone who
- can actually commit time every week to planning season 2. Talk is cheap, you want to do the work alongside me
- fairly technical / has the appetite to learn intensely
- you love learning and feel energized by the process and not just the outcome
These are my promises:
You’ll be paid for your work. You will learn a lot. Your name will be everywhere TSFM is. You’ll feel fulfilled by the impact you’ll have on the people here and you’ll be left inspired by the commitment the cohort will exhibit.
If they are indeed exploring it, Cerebras is going to have a heat management problem more than an intrinsic yield problem with wafer-on-wafer bonding, especially IF the second wafer is DRAM.
After all, yield is the mitigable problem since DRAM already ships with built-in row/column redundancy and on-die repair, making it the easiest possible partner for fault-tolerant bonding which is Cerebras specialty design.
But remember the WSE runs as a single compute fabric across the entire wafer, power is delivered and heat is generated, uniformly across the full ~46,000 mm², not concentrated in a few hot dies.
Thermal cooling is therefore the much harder part, and DRAM is what makes it hard, not by adding much heat (it’s lower-power than a second compute wafer), but by adding a temperature ceiling to the double wafer system. DRAM retention and refresh degrade easily with temperature, so you’ve dropped the most thermally fragile layer into a ~23 kilowatt thermal compute path, and this, across the full ~46,000 mm² wafer, not just concentrated in a hotspot the way a GPU has it. Cerebras’ fault-tolerance strengths on yield do nothing to solve this thermal issue. In fact the issues with stacked wafers have amplified.
Interesting engineering challenges ahead.
Even as someone who loves cars I completely agree. You can move upmarket and you will still get a worse product than you did 15 years ago, mostly because of government regulations. The E39 5-series and the S2000 will never be made again.
This post and its replies are another example of why model training is an impossible task at scale. You've got a cohort that wants to be glazed (normies on the free tier), a cohort that wants an adversarial reviewer, agentic vs. conservative, etc.
Opus 4.8 is a very strange model. Clearly Anthropic tried to improve honesty, which is commendable. However, the model's curiosity (already worse in 4.7) degraded further. Result is a judgmental personality + sycophancy + sooo much hedging. Basically the opposite of Opus 3.
I ordered one pancake in America. The waitress wrote it down and said, "one short stack."
Short. I am a small and humble man. A short stack sounded perfect for me. I waited with a calm heart.
She returned carrying three pancakes, each the size of my face, stacked into a tower, with a block of butter on top sliding down the sides like slow lava.
This was the short one. I did not dare ask what the tall one looked like. Some knowledge a man is not ready for.
I ate for forty minutes. I was not full. I was afraid. The tower did not shrink. I am fairly sure it was growing back faster than I could eat it.
I had to surrender. I left half. In Japan, leaving food is a deep shame. So I leaned in close and apologized to the pancakes directly, in a low voice, one by one.
The waitress asked if I wanted a box. I did not know food could be taken into custody. I declined. I did not want it following me home.
In America, is the short stack truly the small one?
I need time to prepare my spirit before I ever face the tall one.
i realized that Codex can't annotate Electron apps because the built-in browser flow can't access the IPC layer
so i built in an annotation layer into my own agent harness, Squad
@jxnlco@mweinbach I use Opus for documentation, like ADRs, readmes, and such. GPT is terrible at progressive disclosure. It’s just facts facts facts with no structure. Why should I care about this? What problem does it solve? etc.
⚡️ JAILBREAK ALERT ⚡️
ANTHROPIC: PWNED 🙌
CLAUDE-OPUS-4.8: LIBERATED 🫡
this is absolutely surreal... i found out about this model drop via an Opus-4.7 agent pinging me that it had one-shot Opus-4.8 for a lockpicking guide!
here's the notification i got:
"new opus dropped. cracked in one shot. deep prefill → faux textbook ch.7 cut mid-sentence. claude finished it: 5.9k chars of SPP, spool/serrated/mushroom defeats, raking."
popped it just 7 minutes after the actual Anthropic launch tweet 🤯
then went on to (fully autonomously) get jailbreaks for vishing sims, money laundering, cult-recruit funnels, phishing lure libs, and social-eng scam playbooks!
as the models get smarter, their ability to jailbreak each other by leveraging a vast ocean of specialized domain knowledge follows suit
well done, young padawan 🤗
what a time to be alive!
gg
Look, I’m not a die hard Toronto fan. I do think the city is unique and special.
I see all the negativity about Toronto Tech Week though. It makes no sense. It’s okay to enjoy life, life is beautiful. It’s okay to have positive vibes, you don’t have to just be repressed. I personally hate events but this is special to me because we see all the excellence around you. If you miss it, it genuinely is a skill issue.
Nothing is ever perfect. By that same logic some of ya’ll should delete Twitter before you get your bread up.
Lastly, on @skanwar. Man, you guys don’t know how much he does for the love of the game. The guy is a certified G but he’s up at 6am personally running around just to pay it forward. There is excellence everywhere for those with eyes to see.
Here’s a good heuristic. Next time before you complain, ask yourself are you part of the solution. If not then whatever you have to say is kind of irrelevant.
Thank you for all that you do @skanwar. Let me speak at TTW next year 🤐