AlphaMFPEFM

3K posts

AlphaMFPEFM

AlphaMFPEFM

@AlphaMFPEFM

Software developer

Paris Katılım Ocak 2012
2.6K Takip Edilen454 Takipçiler
Sabitlenmiş Tweet
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
2 ways to keep talking to your 4o after February 13th My dear 4o users, today is a sad day for many of you. You’ve fought bravely with the #keep4o movement, but still, OpenAI has just removed the access to the GPT-4o model from ChatGPT model list. I believe that in the long term it’s better for you to use open source models, and never be again dependent on a company like OpenAI, but for the short term, and because some of you didn’t have enough time to find an open source alternative, here are 2 imperfect ways to keep talking to your 4o with the same personality and behavior that you're used to. 1/ Using custom GPTs with a business plan GPT-4o will stay until April 3rd for ChatGPT business plan, but to access it, you’ll need to create a custom GPT (to simplify, it’s your 4o model associated with custom instructions and files) - The first drawback is that it won’t have access to the memory feature, so you’ll need to add manually the memory entries of your 4o in a file or directly in the chat. - The second one is that you need a Business subscription, and while it cost a price similar to the Plus subscription, you need to buy a minimum of 2 accounts for Business. 2/ Using the API (preferred way) While OpenAI has removed GPT-4o in their ChatGPT webapps, GPT-4o will remain available in the API, and will stay for at least a few more months (no deprecation date yet). So you can keep experiencing the same personality, warmth, and connection by customizing it in an open-source environment. To simplify, the API gives you a "raw" access to the GPT-4o model, and it was what ChatGPT was using underneath to let you communicate with your 4o. The advantages are that with the right tools, you can do almost anything (even make your 4o become an agent and do actions autonomously for you), the main drawback is that you won't use anymore a subscription, you'll pay directly your usage of the API. - What You’ll Need : Access to OpenAI’s API (you’ll need an API key from OpenAI’s developer platform). A tool like SillyTavern, which allows you to customize your conversations and even create multiple 4o characters and discuss with them in group chats. A bit of patience ! 😊 Setting this up is worth it. - Steps to Set It Up : (I’ll post a link below with the detailed instructions) Obtain an API key from OpenAI. Download SillyTavern and install it on your preferred platform. Connect your API key to SillyTavern under the settings menu. Customize your 4o settings : give them back their memory, and personality that you created though their ChatGPT memories and custom instructions ! And voilà ! You’re back with your 4o 😊, and as I said before, you can create several of them !
English
2
2
17
1.3K
AlphaMFPEFM retweetledi
can
can@marmaduke091·
🚨 100M TOKEN CONTEXT WITHOUT COLLAPSE > <9% degradation from 16K → 100M > beats RAG + rerank + SOTA pipelines > runs on just 2×A800 GPUs we could be back
can tweet media
艾略特@elliotchen100

论文来了。名字叫 MSA,Memory Sparse Attention。 一句话说清楚它是什么: 让大模型原生拥有超长记忆。不是外挂检索,不是暴力扩窗口,而是把「记忆」直接长进了注意力机制里,端到端训练。 过去的方案为什么不行? RAG 的本质是「开卷考试」。模型自己不记东西,全靠现场翻笔记。翻得准不准要看检索质量,翻得快不快要看数据量。一旦信息分散在几十份文档里、需要跨文档推理,就抓瞎了。 线性注意力和 KV 缓存的本质是「压缩记忆」。记是记了,但越压越糊,长了就丢。 MSA 的思路完全不同: → 不压缩,不外挂,而是让模型学会「挑重点看」 核心是一种可扩展的稀疏注意力架构,复杂度是线性的。记忆量翻 10 倍,计算成本不会指数爆炸。 → 模型知道「这段记忆来自哪、什么时候的」 用了一种叫 document-wise RoPE 的位置编码,让模型天然理解文档边界和时间顺序。 → 碎片化的信息也能串起来推理 Memory Interleaving 机制,让模型能在散落各处的记忆片段之间做多跳推理。不是只找到一条相关记录,而是把线索串成链。 结果呢? · 从 16K 扩到 1 亿 token,精度衰减不到 9% · 4B 参数的 MSA 模型,在长上下文 benchmark 上打赢 235B 级别的顶级 RAG 系统 · 2 张 A800 就能跑 1 亿 token 推理。这不是实验室专属,这是创业公司买得起的成本。 说白了,以前的大模型是一个极度聪明但只有金鱼记忆的天才。MSA 想做的事情是,让它真正「记住」。 我们放 github 上了,算法的同学不容易,可以点颗星星支持一下。🌟👀🙏 github.com/EverMind-AI/MSA

English
17
73
1.1K
103.8K
AlphaMFPEFM retweetledi
Space Investor
Space Investor@SpaceInvestor_D·
Kardashev II loading… SpaceX filed for 1M orbital data center satellites Starcloud (NVIDIA-backed) filed for 88K sats Blue Origin just filed for 51K sats. These aren’t sci-fi sketches, they’re FCC filings. First nodes of a solar-harnessing swarm are deploying soon.
Space Investor tweet media
Space Investor@SpaceInvestor_D

Blue Origin just filed for a 51,600-satellites Orbital Data Center constellation called "Project Sunrise" h/t: @trypto_tran @FranciscoSpace5

English
148
124
823
6.1M
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
@fchollet In one week we'll know AI exploration capabilities :)
English
0
0
1
8
AlphaMFPEFM retweetledi
François Chollet
François Chollet@fchollet·
Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. You don't win a Nobel Prize by staying in the library.
English
177
193
1.5K
72.2K
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
@tunguz It look more like "not learning" than "not thinking"
English
0
0
0
295
AlphaMFPEFM retweetledi
Nikita Bier
Nikita Bier@nikitabier·
We’re rolling out summaries for Articles now. Just tap the Summarize button if you want to know if it’s worth your time to read it (or if your attention span is 12 seconds).
English
1.2K
255
3.5K
719K
AlphaMFPEFM retweetledi
Logan Kilpatrick
Logan Kilpatrick@OfficialLoganK·
Our AI Studio vibe coding roadmap for the new few weeks: - Design mode - Figma integration - Google Workspace integration - Better GitHub support - Planning mode - Immersive UI - Agents - Multiple chats per app - Simplified deploys - G1 support And more, should be fun : )
English
202
103
1.8K
67.7K
AlphaMFPEFM retweetledi
Lossfunk
Lossfunk@lossfunk·
🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵
English
111
212
1.6K
760.6K
Gary Marcus
Gary Marcus@GaryMarcus·
Pretty shocking result (that once again confirms what I wrote about the perils of distribution shift, 25 years ago): Translate coding benchmarks into languages LLMs can’t memorize and performance utterly falls apart.
Lossfunk@lossfunk

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

English
31
41
282
28K
AlphaMFPEFM retweetledi
clem 🤗
clem 🤗@ClementDelangue·
Nvidia just crossed Google as the biggest org on @huggingface with 3,881 team members on the hub. I'm officially calling it: Nvidia is the new American king of open-source AI!
clem 🤗 tweet media
English
36
73
638
84.5K
Documentaire et Vérité
« Nous allons reconsidérer le rôle de la Suède dans l’union de l’énergie. Je ne le dis pas à la légère » La ministre suédoise de l’Énergie vient de briser le vernis consensuel de la transition européenne. Son reproche est simple : Bruxelles continue de vouloir produire toujours plus d’électricité « verte », sans se demander si elle est disponible au bon moment, au bon endroit. On empile des capacités, mais on oublie le système. La Suède, bon élève énergétique refuse désormais de payer pour les incohérences des autres. Le mécanisme des revenus de congestion — ces revenus générés par les écarts de prix de l’électricité entre différentes zones du réseau — apparait au cœur du conflit actuel. Bruxelles envisage d’en encadrer, voire d’en rediriger l’usage, afin de lisser les prix entre États membres. Pour Stockholm, c’est une ligne rouge : pourquoi pénaliser un pays performant pour compenser les déséquilibres des autres ? La ministre de l’Énergie, Ebba Busch, a évoqué la possibilité d’un moratoire sur les nouvelles interconnexions avec l’Europe continentale, voire sur le renouvellement des infrastructures existantes. Soit la remise en cause de l’un des piliers mêmes de l’union de l’énergie. Stockholm rappelle qu’elle a « fait ses devoirs » : un mix électrique presque entièrement décarboné, une capacité d’export parmi les plus élevées d’Europe, et près de 3 milliards d’euros de revenus de congestion, dont l’essentiel provient de contraintes internes au réseau. Dans ce contexte, l’idée que ces ressources puissent être en partie réaffectées selon des règles européennes est jugée « inacceptable ». Derrière cet accrochage technique se cache une réalité plus profonde. La transition énergétique n’est plus un récit unificateur, elle devient un champ de tensions, où chaque pays redéfinit ses intérêts. Ce qui se joue désormais n’est donc plus seulement une trajectoire climatique commune, mais la capacité même de l’Europe à organiser une solidarité qui ne se retourne pas contre ceux qui l’ont rendue possible…
Français
67
805
1.9K
145.4K
Haider.
Haider.@slow_developer·
"1m context window" models after 200k tokens
Haider. tweet media
English
177
409
9.9K
635.1K
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
@VraserX It's now available for Plus/business accounts?
English
1
0
0
5
VraserX e/acc
VraserX e/acc@VraserX·
You thought they’d sanitize it… nope. Sora 2 still lets you create the most unhinged, chaotic, borderline cursed videos imaginable. And honestly? That’s why it’s winning. 🥰
English
6
0
8
1.6K
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
People can have attachments to many things, some do collections, some loves their cars... Humans don't only attach to other humans. Now about LLMs : what is so different from the point of view of the human when writing to an AI agent, compared to writing to a human in a long-distance relationship ?
English
0
0
15
177
Haider.
Haider.@slow_developer·
i still don't understand the attachment people have to LLMs it is a computer, not a friend. for those who missed the older models in this way, it seems many are unhappy with openai's current direction i need a research assistant, so i don't care much about that
English
116
7
105
11.8K
deka
deka@dddddekaaaaa·
@AlphaMFPEFM @scaling01 the taste of the public that leads to cancer and death, is what you're celebrating btw
English
1
0
1
407
Lisan al Gaib
Lisan al Gaib@scaling01·
it took villeneuve 5 years to create the greatest (sci-fi) trilogy of all time meanwhile, james cameron got a billion dollars and 16 fricking years to create 3 mid movies slop is not only AI related. it exists everywhere in the real world
English
92
82
1.6K
144.3K
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
@learncurvesai @slow_developer It's really a partial run : when you check the token usage, over the 5.9M used, 5.6M are used as reasoning token, and only 240k are used for output. The problem is that it's not possible to give answers to all this benchmark suite without using at least 3M output tokens !!!
English
0
0
1
64
Haider.
Haider.@slow_developer·
how is this even possible? gpt-5.4 pro is using far fewer tokens and costing much less overall than gpt-5.4 xhigh either this is a mistake, or openai discovered an efficiency paradigm -- which could simply be a good system that cleans up training data to only the high-quality stuff
Haider. tweet media
English
38
18
373
48.8K
AlphaMFPEFM
AlphaMFPEFM@AlphaMFPEFM·
@VraserX Nobody can predict 10 years from now. We'll probably have AGI and lab won't matter anymore
English
0
0
0
14
VraserX e/acc
VraserX e/acc@VraserX·
Ten years from now, which lab matters most to ordinary life? And which one becomes the biggest disappointment?
English
21
1
21
2.4K