deko retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
be careful with @Muse , seems like it very non optimized for the risks of using too much quota at once. I did same ask for it claude and codex to read some old claude code session and summarize the content for me, outcome, claude and gpt didnt spent as much quota, claude was going to spawn subagents to read whole conversation but I stopped it before but I believe wouldnt have spent whole quota
muse just spawn bunch of sugbagents and burn my muse developer power usage quota in a minute to read the whole covnersation, completely not want I want or would have approved.
It didnt even send a message, it just went directly to spawn subagents and burn it all. Usage data showing " Subagents 18 completed (6,177,203 tokens)" and this is in a new session without any message sent from muse to me, this is not good behavior, you should improve this.
Lmk if you need more data, would be happy to contribute, htis was via muse code, maybe I should use it via claude code/codex/opencode could give better results of an expected mature system
@finkd @alexandr_wang
deko retweeted
We are making Markdown files work natively across Google Drive and Docs.
You can now preview .md files in Drive, open them in Docs, edit them, comment on them, and collaborate without converting the file into a Google Doc!
deko retweeted
Japan's biggest power generator is now in the compute business.
JERA just signed with US's @Dell and UK's RHAELM to build the largest single-site AI deployment in Asia outside China.
$15B. 400MW. One site in Chiba (next to Tokyo).
I've been saying it. Come build with Japan!
In Japan we're putting the AI factory next to the power plant.
JERA + RHAELM + @Dell + @apolloglobal: 400MW in Chiba, first phase online in 2028.
It will be the largest single-site AI deployment in Asia outside China.
Multi-gigawatt across Japan next. 🇯🇵
rhaelm.com/news/jera-dell-rh…
deko retweeted
I don’t blame OpenAI for serving much more powerful versions through the API while limiting them in ChatGPT / Codex.
If they didn’t, they’d probably burn through insane amounts of compute. APIs are expensive for a reason, and now that reason makes a lot more sense.
What I DO criticize is the lack of transparency.
Are the benchmarks we see measuring ChatGPT or API performance? How much does model quality change between the API and ChatGPT? Is “Max” actually the same in both environments?
Those differences matter, and users deserve clearer answers.
Devin @devindesktop cli is good and model is good, but is crazy memory hungry and uses even more to start. I can see my memory going up a few GB and them down every time I start one new session.
It crashed my herdr quite a few times because I have many sessions open and when it starts everything it just freeze the pc and crashes everything as it uses whole system memory
I have @AnthropicAI but Opus 5.5 is freaking amazing, they really got back to the game after just producing too expensive models.
This is like the perfect model for current AI stage
Niceee! Thats actually something we really need. Its pretty obvious for power users whats going on
deko retweeted
🚨 Japan is About to Nuke the Global Economy
A very dangerous news came out from Japan for Carry Traders: BoJ’s Yuto just revealed that Important off-the-record deals have been struck to secure Japan’s oil demand.
Markets immediately heard the implication: oil priced in yen. If that happens, the decades-old yen carry trade does not just wobble. It snaps.
Oil priced in yen would mean Japan (or its counterparties) settling at least some crude purchases in yen instead of dollars. Oil has been a dollar market for decades. Japan earns in yen, so it normally sells yen for dollars, then pays for barrels.
That constant yen-selling is one of the structural pressures that kept the currency weak and made cheap yen funding attractive. If a growing share of Japan’s oil bill is paid in yen, especially under bilateral political deals with the United States after the Hormuz disruption, two things change at once:
Japan needs fewer dollars for energy, and counterparties who accept yen have a reason to hold or recycle yen. That reduces a major source of yen selling.
That is what hit markets in August 2024. At the same time, Japanese yields are rising and yen is weakening after multiple interventions by U.S. Treasury and Fed. The U.S. 10-Year Yields have already broken 5.18% and not seeming to stop.
Scott Bessent even dumped Euros to save the yen, doubled U.S. Bond buybacks, warned the Fed to expand FIMA facility to Japan or watch the U.S. Treasuries collapse.
A move toward yen-denominated oil would reinforce the yen, raise the cost of funding the old carry, and accelerate the shift toward the reverse trade.
Higher yields in both countries already tighten financial conditions; a sudden yen squeeze on top of that would force leveraged positions to unwind into rising U.S. and Japanese bond yields at the same time.
This was exactly warned by the Famous City of London banker @LordBelgrave at the start of the year.
The Yen-Crude Settlement was inevitable. Looks like it’s finally arriving.
Lol my codex has just reset and I had 99% of quota and them @thsottiaux and @OpenAI reset it again. Thats ridiculous
Maybe I will not renew the 200 plan, will get 2x100 and keep them in different reset dates as this is just stupid
Do not use @StepFun_ai with your data before reviewing their data policy. Really really bad and worrying
Review from astra: chatgpt.com/share/6aafaff8-3…
Also cause I find astra the worse model from got in a long time. Ofc its crazy smart but uses token incredibly highly, absolutely what we hated from claude and we fell in love with got cause it was smart and usable without destroying quota. It did so great on 5.5 and 5.6 and them they do the same thing than claude
This seems pretty great, maybe I should subscribe to Claude again
deko retweeted
Extraordinary claims require extraordinary evidence so check out our release blog for more technical info: typesafe.ai/blog/introducing…
Join our waitlist for early access: typesafe.ai/
Have technical chats and meme with us on Discord (rumors are good memers skip the line): discord.gg/WUujKYBp8s
We are so pumped this is in the hands of developers now, and we are just getting started!!!
Im starting to hate Astra, you tell it to do something and it keeps doing it for hours just to find out was completely wrong.
I have max and gave a task for doing some big work, 50% of usage later, absolutely went out of what it should deliver and deliver something completely wrong
Last time similarly, added a bunch of bullshit to it that I had to refactor for a while
It just get what you say and go on and start working until it gets whatever it feels it should deliver. Feels like @OpenAI is just making it easier to stick with sessions and work so it can spend more tokens faster












