اکنون در دسترس! پژوهش تلگرام ۲۰۲۵ — مهمترین بینشهای سال 
Biniyam
رفتن به کانال در Telegram
Biniyam here 👋 I build AI products and this is a channel that I share my journey
نمایش بیشتر1 361
مشترکین
-224 ساعت
+207 روز
+20830 روز
در حال بارگیری داده...
کانالهای مشابه
هیچ دادهای
مشکلی وجود دارد؟ لطفاً صفحه را تازه کنید یا با مدیر پشتیبانی ما تماس بگیرید.
ابر برچسبها
اشارات ورودی و خروجی
---
---
---
---
---
---
جذب مشترکین
ژوئیه '26
ژوئیه '26
+1
در 0 کانالها
ژوئن '26
+230
در 31 کانالها
Get PRO
مه '26
+106
در 12 کانالها
Get PRO
آوریل '26
+55
در 2 کانالها
Get PRO
مارس '26
+177
در 12 کانالها
Get PRO
فوریه '26
+148
در 10 کانالها
Get PRO
ژانویه '26
+422
در 44 کانالها
Get PRO
دسامبر '25
+149
در 11 کانالها
Get PRO
نوامبر '25
+187
در 0 کانالها
Get PRO
اکتبر '250
در 7 کانالها
Get PRO
سپتامبر '250
در 0 کانالها
Get PRO
اوت '250
در 2 کانالها
Get PRO
ژوئیه '25
+9
در 5 کانالها
| تاریخ | رشد مشترکین | اشارات | کانالها | |
| 01 ژوئیه | +1 |
پستهای کانال
Crazy part is Meta only used ~2k typed sentences for v1, then 22k for v2 (10x more data).
Character accuracy: v1 61% → v2 69%
Word accuracy (best participant): v1 48% → v2 78%
This shows with a scaled-up dataset you can get way more accuracy. Huge leap in tech, especially for people with disabilities.
@b1n1yamBuilds
| 2 |  Meta dropped Brain2Qwerty today.
It turns raw brain signals into real-time sentences. You’ll need a MEG helmet but this was only possible with surgery before.
v2 hits ~61% word accuracy on average, up to 78% for the best person, decoding full thoughts into text (TTT 😄)
They open-sourced the code too. Wild how fast this is moving.
@b1n1yamBuilds | 396 |
| 3 |  So this is the new normal.
OpenAI announced its new suite of 5.6 models, but it’s not going to be generally available yet. Surprise surprise.
And yes, the U.S. government is now involved in who gets early access.
The inequality with AI is going to be like no other.
#b1n1yamBuilds | 1 339 |
| 4 | Sanyi is back with a new channel. Go sub! | 1 083 |
| 5 | I just learned that the thing we know as ደረቅ / ደረቅ የሆነ ሰው actually has an old Greek philosophical word.
It’s called Amathia.
The closest definition is: the active refusal to know.
It’s ignorance, but on a whole new level.
With ordinary ignorance, you can sometimes fix it by giving someone a book, a documentary, an experience, or better information.
But Amathia is way darker.
It’s the student who thinks they are too good for their teacher’s feedback.
It’s when a person refuses to learn anything that might challenge what they already believe.
The Greeks saw Amathia as one of the greatest causes of human evil.
Because if someone walks around with the self-conviction that they already know enough, they stop trying to learn more.
And if someone thinks they are already perfect, they stop trying to grow.
@b1n1yamBuilds | 1 698 |
| 6 | Tech should meet where people are | 1 703 |
| 7 | For the non readers out there, here is a voice over generated by Addis Voice 2
@b1n1yamBuilds | 1 966 |
| 8 | https://x.com/satyanadella/status/2066182223213293753?s=46
“You can offload a task, or even a job, but you can never offload your learning”
Satya just said something every founder needs to hear.
Everyone is obsessed with one question: which AI model is the best?
And he says that’s the wrong question to ask
The model is not your advantage, the advantage is the learning loop you build on top of it. He says there are two capitals
Human capital: your people’s judgment, taste, relationships, and instincts.
Token capital: the AI system your company owns and trains. The agents and harness around it
So when your people do the work through that harness and the work improves the AI. It scope it and teaches it what’s important
And private evals and RL environments will identify traces of success in the models performance. Then slowly it will be come proficient in your brand, vision, priorities and problems not through prompts but through training.
Any model that comes afterwards (could be the strongest model out there ) wouldn’t out perform your framework as it’s grounded in your context.
So in a way this loop becomes your IP
The whole post is a great read, but I’d recommend using TTS to listen to it out loud.
@b1n1yamBuilds | 1 615 |
| 9 | This is the US straight up telling the world their stance. Even if they build AGI (btw trained on dataset across the globe) they’re most likely not sharing it with us.
የሰው ነገር የሰው ነው😏
That’s why we need our own models. More African/Ethiopian startups building AI infra. Depending on gatekeepers is always gonna put us in a disadvantage.
@b1n1yamBuilds | 3 100 |
| 10 |  The thing is Anthropic has no reliable way to verify users’ nationality “yet”, so they had to disable Fable 5 for everyone to comply
@b1n1yamBuilds | 1 283 |
| 11 |  We might be seeing GPT-5.5 getting the same fate as Fable 5.
Anthropic is playing the old "አኔ ከሞትኩ ሰርዶ አይብቀል" game.
@b1n1yamBuilds | 1 048 |
| 12 | This's the second time it has happened in a few months. First was @FrostySiren now @TheSanyi — it's so annoying and their support team doesn't even respond. | 866 |
| 13 |  Altman rn | 1 043 |
| 14 |  Well that didn't last long.
Fable 5 just got hit with an immediate suspension from the US gov.
Maybe, just maybe… don’t advertise your model as “too dangerous” and then drop it to the public the next week.
@b1n1yamBuilds | 1 897 |
| 15 |  This is how a company that listens to its users and actually cares about them behaves | 1 105 |
| 16 |  I don’t blame you, YouTube.
Good coffee + 😍 is the way to a productive day. | 959 |
| 17 |  We’ll all be saying “ni hao” soon 🙏 | 1 737 |
| 18 | I really can’t wait for the day where people build things just by speaking in their own language.
They don’t have to know code or even English for that matter. Just their own natural language.
Like someone in a rural area, or someone who has never written a line of code, just explaining (through voice) a problem to their phone, then watching it turn into a real product.
There are thousands of small problems around us that never get solved because the people closest to them don’t have the tools or access to build.
I know this future is coming. I just really hope it is not far.
@b1n1yamBuilds | 1 133 |
| 19 |  Okay this is impressive.
I gave Fable 5 access to our model playground, previous runs, scripts, datasets, and benchmarks from earlier models, then asked it to improve our fine-tuning pipeline.
And frens, it actually did!
It ran 3 experiments on instruction tuning our already CPT-trained model, then found a setup that reached 97% of the full run’s Q&A benchmark score, but with around 30% less fine-tuning budget.
Out of those 3, it landed on the result with fewer LoRA config sweeps, fewer training steps, and a better choice of LR scheduler / LoRA rank, while still maintaining quality.
@b1n1yamBuilds | 1 056 |
| 20 |  Here is how we implement loops: openClaw as the harness, hooked to a stronger model as the orchestrator. In our case, we’re using Opus for that layer. | 1 088 |
