sohail

inference perf @togethercompute | ex @amazon @wendys
kernels, gpus, intelligence

[ClaudeDevs @ClaudeDevs]

We've reset 5-hour and weekly rate limits for all users.

[Tejal Patwardhan @tejalpatwardhan]

GPT-5.6 sol post-trained luna!

[sohail @sohailmo]

api to the physical world is such a good frame

[dar @radbackwards]

Alien Technology

[sohail @sohailmo]

using less tokens (cheaper) to get the same level of work done (if not better) is as much of a metric of intelligence as the end score is

comes back to the idea of being able to communicate an idea in as few words as possible as a marker of deep understanding

[sohail @sohailmo]

the 5.6 blogpost is v satisfying to read, lots of interactive bits to let you see the outputs yourself, v curious to test out the long context usecases - cautiously optimistic but scores showcase higher than opus rn (also abstract reasoning seems much better)

[sohail @sohailmo]

'looks like im finally allowed to talk about it...'

[sohail @sohailmo]

@matt_slotnick every post i see about sol sums to it being a diligent worker you can trust and fable being the insightful/intelligent lead, feels like it tbh

[sohail @sohailmo]

great read for anyone curious about agentic inference optimization

https://t.co/0hiprcDinn

[sohail @sohailmo]

cats will never be replaced by ai

[sohail @sohailmo]

with hardship comes ease

[Rajjath @Rajjath24]

it is an unwritten rule of life that after every prolonged period of hardship and uncertainty, there is going to be a period when you are going to achieve quantum leaps across multiple areas of your life. the only requirement is that you do not give up on yourself.

[sohail @sohailmo]

@shensi @awscloud have been very bullish on the need for a strong router, love the work yall are doing keep it up

[sohail @sohailmo]

if you’re bored learn about speculative decoding its v fun when you dive into it

[sohail @sohailmo]

biased as an fde but i think it’s heavily domain dependent

fde worked for palantir bc you literally have to work embedded for highly confidential, critical, and custom systems

solution architects at together for example cover what most fdes at other companies do

seeking value creation for the end user of a product in as bespoke of a manner as possible is what it comes down to

[sohail @sohailmo]

for anyone who’s in engineering and feels burnt out

i’ve recently felt like moving closer to the customer makes a huge difference in how motivated you feel

exciting domain is a big part of it but knowing your work directly affects another persons success is fuel enough

[sohail @sohailmo]

love that when thinking machines dropped their duplex arch, it was applauded and when openai drops the same thing but with an actually usable interface, the upset bros surface

yall are sheep man

[sohail @sohailmo]

@OpenAI @jxnlco @jxnlco the switch in taste of the ads is ~chefs kiss~

[sohail @sohailmo]

highly suggest anyone running production grade inference engines to deeply understand the tradeoffs of jetspec vs dspark

tldr: dspark is optimized for high concurrency/large batch serving systems whereas jetspec wins in low concurrency regimes, these are engine level changes

[Hao AI Lab @haoailab]

https://t.co/lwO6dSHm9m

[sohail @sohailmo]

@NVIDIAAIInfra @togethercompute @realDanFu that’s my VP!!!

[Together AI @togethercompute]

We're introducing Provisioned Throughput: reserved inference capacity for frontier open models, with token-based pricing and a 99% uptime SLA.

Serverless simplicity, guaranteed capacity, up to 90% lower cost vs. Opus 4.8.

Get started with MiniMax M3 + GLM-5.2, read more 🧵

[sohail @sohailmo]

and there you have it folks

[Together AI @togethercompute]

GLM 5.2 on Together AI is now #1 on Artificial Analysis for both output speed and latency.