Pinned Tweet
So, don't take my opinion on LaunchApp Animus CLI.
This is what grok says, if that sounds like something you want repo in my profile. If you want to get stuck in the AI slow lane, then do whatever it is you're doing.
This is what grok says, if that sounds like something you want repo in my profile. If you want to get stuck in the AI slow lane, then do whatever it is you're doing.
2
0
14
1,265
Interesting perspective....
🚨🚨🚨 A warning to Americans, from a Frenchman. 🚨🚨🚨
You are starting to elect people who promise you more state, more redistribution, more "free" things. New York already has a democratic socialist mayor, perhaps the furthest-left since Fiorello La Guardia. I'm not here to 'm here to tell you how this story ends, because we lived it. France ran this experiment for 40 years, and we are now living its final chapter.
It starts in May 1981. François Mitterrand wins the presidency with the Communists in government. The program is pure romance. They nationalize the big banks and industrial groups. They lower the retirement age to 60, cut the work week, add a fifth week of paid vacation, create a wealth tax, and hire massively in the public sector. "Changer la vie", change life itself. That was the slogan.
Reality took less than two years to answer. Capital fled. The franc was devalued three times. Inflation stayed in double digits while our neighbors were bringing theirs down. In March 1983, Mitterrand made a U-turn and imposed austerity, "le tournant de la rigueur".
Remember this, because it is the pattern. Socialism doesn't fail once and disappear. It fails, reinvents itself, and changes its electorate.
After 1983 the Socialists had a problem. They had abandoned the economic revolution, so what was left to sell? The answer was identity politics before the term existed. In 1984 SOS Racisme was born, closely tied to the Socialist Party. The moral battle replaced the economic one. Anyone who raised the question of immigration or integration was no longer an opponent. He was a racist.
Then came the first real test. In 1989, three girls showed up to a public school in Creil wearing the Islamic veil. The Socialist government refused to decide, sent the question to the courts, and let it rot for fifteen years. The ban only came in 2004. That was the moment the Republic blinked. The message received was simple: if you push, the state will back down.
Over the following decades, the left slowly replaced the working class it had lost with a new coalition. In 2011 the Socialist think tank Terra Nova wrote it down in black and white, saying they should stop chasing blue-collar workers and build a coalition of graduates, youth, and minorities. Electoral clientelism became strategy. In certain cities, local politicians traded subsidies, jobs, and silence for votes delivered by community networks, including Islamist ones.
The result today is that in the 2022 presidential election, Muslim voters gave 69% of their first-round votes to Jean-Luc Mélenchon, the hard-left candidate. A party that once claimed to be secular is now courting religious identity. Let's be precise: this is not about our Muslim fellow citizens, millions of whom are fully French. It is about a political machine that decided a community vote was worth more than the principles of the Republic.
Now the economy, because the two collapses are the same collapse.
In 2000, the Socialist government of Lionel Jospin, with Martine Aubry as labor minister, imposed the 35-hour work week. The theory was that sharing work creates jobs. What it actually did was make French labor more expensive, destroy the culture of effort, disorganize hospitals that still haven't recovered, and teach an entire generation that work is a burden the state should shrink for you.
That is the deep logic of socialism. It doesn't make people richer. It makes them dependent. Every new subsidy creates a voter who fears losing it. Every new public job creates a voter who defends the system. After a while, you no longer have citizens. You have clients.
Look where it ends. France hasn't voted a balanced budget since 1974. Public debt was around 20% of GDP when Mitterrand arrived. It is now around 3,550 billion euros, about 118% of GDP. France is once again Europe's biggest spender, with public spending at 57.3% of GDP. And a report commissioned by the government in July 2026 projects that without budget efforts, debt would reach 130% of GDP by 2030.
And to be completely honest, the right didn't fix any of this. Chirac, Sarkozy, Hollande, Macron: they all managed the decline instead of reversing it. The left set the direction, and the right lacked the courage to change it. That's how you get 40 years of slow rot, economic and social at the same time. A country that works less, pays more, gets less security, and no longer dares to say what it is.
So today France has two choices, and so will you if you go down this road.
The first choice is to stop lying to ourselves. Multiculturalism, meaning many cultures living side by side with no common rules, does not work. It never has, anywhere. When there is no shared culture on top, one culture eventually takes over the others. That's not hate, it's sociology. What worked in France for 150 years was assimilation. You come, you become French, you take on the language, the laws, the history, the way of life. In exchange, the nation takes you in fully, with no asterisk. Assimilation plus controlled immigration. And the deal is already signed with millions of people. Children of Italian, Portuguese, Polish, Algerian, Moroccan, and Vietnamese immigrants who are as French as I am. They are not the problem. They are the proof that the model works when we have the courage to apply it.
The second choice is to keep pretending. Let the state get bigger, let debt explode, keep buying social peace with money we don't have, and keep handing territory to those who reject the West. In that scenario, it is not impossible that France first elects a hard-left government, a Mélenchon government. Then, when the system can't pay its promises anymore, something else fills the vacuum. Michel Houellebecq wrote that story as fiction in 2015, in a novel called Submission. A lot of French people stopped laughing at it a long time ago.
France is a Western country with Western values. Equality between men and women. Freedom of expression, including the freedom to criticize religion. The law above faith. Work as dignity, not punishment. Some things are simply not acceptable here, and we need to have the courage to say it again.
America, you still have what we're losing: a culture of work, a culture of assimilation, the idea that the state is your servant and not your parent. The people selling you "more state" are making the same promise Mitterrand made in 1981. It sounds generous. It ends in debt, dependence, and fragmentation.
Don't trade the American Dream for the French Disease. We already paid the price. Learn from it for free.
0
0
0
12
Proving the point that safeguards are an illusion. The only safety is red teaming your own shit.
Period. No other options now. If you've got high visibility, high impact software you better be pen testing it.
We used to have be pentested quarterly back in my payment days.
Period. No other options now. If you've got high visibility, high impact software you better be pen testing it.
We used to have be pentested quarterly back in my payment days.
I strongly believe a part of why so many critical vulnerabilities are happening all over the proliferation of "abliterated" models, or models with safeguards removed.
You can access an abliterated GLM 5.3, a near frontier AI model, on @abliteration_ai today. And it's a world r for like $10.
0
0
0
20
Rafael Aleksandryan retweeted
BREAKING: AI token spending grew just +0.2% MoM in September, the lowest monthly growth since November 2025.
This follows a +1.0% increase in August, marking two consecutive months of growth at or below +1.0%, the weakest streak in at least 18 months.
This comes as falling anced AI models, alongside improvements in efficiency and quality, drive down the average price per million tokens processed by these models.
Since July, overall AI token prices have declined -48%, to $0.65 per million tokens, the lowest since at least January 2026.
Lower costs could unlock the next wave of AI adoption.
This follows a +1.0% increase in August, marking two consecutive months of growth at or below +1.0%, the weakest streak in at least 18 months.
This comes as falling anced AI models, alongside improvements in efficiency and quality, drive down the average price per million tokens processed by these models.
Since July, overall AI token prices have declined -48%, to $0.65 per million tokens, the lowest since at least January 2026.
Lower costs could unlock the next wave of AI adoption.
189
181
1,524
232,696
Rafael Aleksandryan retweeted
if you're using claude to edit videos, give it 17M+ films and 14M audio recordings for $0
the internet archive has a free api. no key.
10,000+ prelinger films (public domain stock footage)
187,000 78rpm records
300,000+ live concert recordings
give it archive.org/developers/ need archival b-roll or period music, search the prelinger and 78rpm collections on archive.org and use them in the edit”
some great clips that are amazing for slightly more artsy ads for tech
the internet archive has a free api. no key.
10,000+ prelinger films (public domain stock footage)
187,000 78rpm records
300,000+ live concert recordings
give it archive.org/developers/ need archival b-roll or period music, search the prelinger and 78rpm collections on archive.org and use them in the edit”
some great clips that are amazing for slightly more artsy ads for tech
67
206
3,547
228,091
Open weight. Stronger than Fable 5 AND cheap (ish) at $10/output not cached.
It's a bit thirsty on tokens, but impressive metrics regardless.
Deepseek is still the stand out because of its efficiency though!
It's a bit thirsty on tokens, but impressive metrics regardless.
Deepseek is still the stand out because of its efficiency though!
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, ude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
0
0
0
29
Rafael Aleksandryan retweeted
Today we announce 1Password for Agents platform, building with companies @Muse, SpaceX Super Intelligence Grok @bot, @AnthropicAI’s @claudeai, @OpenAI, @TownAI, @browserbase, @browser_use, @heydana_ai, Instinct, @usekernel, @perplexity_ai, @tryramp, @tomo, and @wajo_ai, with e way. 🎉
We've heard from so many of our customers who want to use these new personal AI agents to help them get more done, but want more control over what the agent can access with the same trust and ease of use they get when using 1Password directly.
1Password for Agents gives your agents access without giving up your passwords. When an agent reaches a sign-in form, it requests a credential from your vault. If you approve the request, 1Password directly auto-fills the credential in the destination site so the agent can’t see or take possession of the credentials.
The agent gets the task done, but it doesn’t get your passwords..
You stay in control. 1Password for Agents free and optional for personal and family accounts. Unless you opt-in to an agent connector, agents cannot touch your vault.
➡️Learn more: bit.ly/487da6J
We've heard from so many of our customers who want to use these new personal AI agents to help them get more done, but want more control over what the agent can access with the same trust and ease of use they get when using 1Password directly.
1Password for Agents gives your agents access without giving up your passwords. When an agent reaches a sign-in form, it requests a credential from your vault. If you approve the request, 1Password directly auto-fills the credential in the destination site so the agent can’t see or take possession of the credentials.
The agent gets the task done, but it doesn’t get your passwords..
You stay in control. 1Password for Agents free and optional for personal and family accounts. Unless you opt-in to an agent connector, agents cannot touch your vault.
➡️Learn more: bit.ly/487da6J
630
1,075
10,499
1,732,782
Rafael Aleksandryan retweeted
The good news: you can mid-train yourself to the frontier!
At a fraction of the cost! Without distillation!
What we did: We took Qwen3.5 14B and mid- and post-trained it to beat every other open model and perform as well as the very best frontier models.
This is no simple ng, we are training the model to understand and reason over DNA sequences. Bio is the next frontier for AI and universal access will be crucial to make progress.
The fun part: because we train the model on both natural language and DNA sequences you can actually 'chat to your DNA': Ask it questions, ponder the impact of mutations, or ask it to generate a small sequence.
We'll soon see the frontier labs move away from math and cyber towards bio. It already started, and we believe researchers should have access to capable models.
We are releasing the full recipe, so you can do it, too!
Model: huggingface.co/HuggingFaceBio…
Blog: huggingface.co/spaces/Hugging…
At a fraction of the cost! Without distillation!
What we did: We took Qwen3.5 14B and mid- and post-trained it to beat every other open model and perform as well as the very best frontier models.
This is no simple ng, we are training the model to understand and reason over DNA sequences. Bio is the next frontier for AI and universal access will be crucial to make progress.
The fun part: because we train the model on both natural language and DNA sequences you can actually 'chat to your DNA': Ask it questions, ponder the impact of mutations, or ask it to generate a small sequence.
We'll soon see the frontier labs move away from math and cyber towards bio. It already started, and we believe researchers should have access to capable models.
We are releasing the full recipe, so you can do it, too!
Model: huggingface.co/HuggingFaceBio…
Blog: huggingface.co/spaces/Hugging…
37
124
1,307
78,536
Rafael Aleksandryan retweeted
The TSO, or Technology Support Office, was established in April to 'covertly disrupt and sabotage foreign technology programs considered threats to American military superiority.'
In September the Washington Sun reported that the NSA told lawmakers in classified intelligence ates it has spent billions of dollars 'evaluating and testing' models this year already.
In the draft S-1 leaked by the WSJ yesterday, Anthropic said that one-fifth of its $16.3 billion in revenue in the first half of 2026 came from just two customers.
Motive, means, and opportunity. I think a lot of money is being spent by agencies through Project Glasswing or channels related to it, and they have probably covertly penetrated every nation on earth to look at anything even potentially considered a threat. There is no defense against the current internal version of Mythos unless you have access to it. No foreign intelligence agency on earth has anything even close. It's actions would also almost certainly be undetectable. Assume US agencies know much more about any technological progress in China than they are letting on; given Dario's openly stated position on China, this would almost certainly be done with his approval.
In September the Washington Sun reported that the NSA told lawmakers in classified intelligence ates it has spent billions of dollars 'evaluating and testing' models this year already.
In the draft S-1 leaked by the WSJ yesterday, Anthropic said that one-fifth of its $16.3 billion in revenue in the first half of 2026 came from just two customers.
Motive, means, and opportunity. I think a lot of money is being spent by agencies through Project Glasswing or channels related to it, and they have probably covertly penetrated every nation on earth to look at anything even potentially considered a threat. There is no defense against the current internal version of Mythos unless you have access to it. No foreign intelligence agency on earth has anything even close. It's actions would also almost certainly be undetectable. Assume US agencies know much more about any technological progress in China than they are letting on; given Dario's openly stated position on China, this would almost certainly be done with his approval.
48
151
1,058
89,682
Rafael Aleksandryan retweeted
Scientists found a way to flush Alzheimer’s plaque out of the brain in minutes.
Without any surgery and drugs.
Your brain has a waste-clearance network called the glymphatic system. It flushes out toxic proteins and dead cells.
But it only really activates during deep, u don't sleep, the trash piles up. Over time, that buildup leads to cognitive decline and Alzheimer's.
Until today.
Researchers just published a paper proving they can manually trigger this cleaning cycle.
Without you ever going to sleep.
They used 40.68 MHz ultrashort waves (USW) beamed non-invasively into the brain.
Within 10 minutes, the brain's fluid flow rapidly accelerated.
The waves artificially synchronized the neurons into a Delta state, the exact frequency of deep sleep. This forced the brain's water channels to open and start flushing out the system.
They tested it on mice with traumatic brain injuries and Alzheimer's pathology.
The brain swelling dropped. The toxic proteins were cleared out. Cognitive deficits actually improved.
No surgery. No massive chemical interventions. Just a frequency.
Without any surgery and drugs.
Your brain has a waste-clearance network called the glymphatic system. It flushes out toxic proteins and dead cells.
But it only really activates during deep, u don't sleep, the trash piles up. Over time, that buildup leads to cognitive decline and Alzheimer's.
Until today.
Researchers just published a paper proving they can manually trigger this cleaning cycle.
Without you ever going to sleep.
They used 40.68 MHz ultrashort waves (USW) beamed non-invasively into the brain.
Within 10 minutes, the brain's fluid flow rapidly accelerated.
The waves artificially synchronized the neurons into a Delta state, the exact frequency of deep sleep. This forced the brain's water channels to open and start flushing out the system.
They tested it on mice with traumatic brain injuries and Alzheimer's pathology.
The brain swelling dropped. The toxic proteins were cleared out. Cognitive deficits actually improved.
No surgery. No massive chemical interventions. Just a frequency.
602
7,671
32,012
1,431,275
So...who's got that OpenSource Android TV...
WTF, $3000 Samsung TV has an always-on mic, automatic content recognition, geo tracking, network mapping, and ships your data to Samsung + 400 partners.
Bro Rooted a brand-new 2025 Samsung TV after noticing absurdly targeted ads.
Firmware was locked down, so zero-days it was. n found: always-on voice service, ACR, geo, full telemetry pipeline, and four separate DRM systems.
This isn’t a bug, It’s the product bro.
0
0
0
13
Pretty good, but Chinese models are still goated.
Mistral has released new Speech models! Voxtral Small Transcribe 3 takes the #3 spot on the Artificial Analysis Speech to Text leaderboard, and Voxtral 3 TTS enters the top 12 in both Text to Speech arenas
@MistralAI has released Voxtral 3: a new Text to Speech model and three Small Transcribe 3, Mini Transcribe 3 and Mini Transcribe Realtime 3). We evaluated all of them.
Key takeaways:
➤ Speech to Text: Voxtral Small Transcribe 3 takes the #3 spot on the Artificial Analysis Speech to Text leaderboard with a 2.1% AA-WER Index, ahead of ElevenLabs Scribe v2 (2.2%) and Grok Voice Transcribe 2.0 (2.3%). Voxtral Mini Transcribe 3 lands at #5 with 2.2%.
➤ Streaming Speech to Text: Voxtral Mini Transcribe Realtime 3 scores 3.3% final-transcript AA-WER at the 960 ms setting (#6 of 46) and 3.6% at 480 ms, down from 5.2% for the previous Voxtral Mini Transcribe Realtime.
➤ Text to Speech, Provider Voice: Voxtral 3 TTS scores an Elo of 1,218 across 1,599 arena appearances, ranking #12 of 98, just behind Breeze TTS 2 at 1,221 and ahead of Inworld Realtime TTS-2 Flash and StepAudio 2.5 TTS at 1,216. It ranks #13 in English (US) and #12 in English (UK).
➤ Text to Speech, Controlled Voice: Voxtral 3 TTS scores 1,068 across 1,614 appearances, ranking #12 of 43 with the same 8 cloned voices, behind Eleven v3 at 1,072 and level with Ant Group's Ming-3.0-tts at 1,068. It ranks #10 in Assistants and #10 in English (UK).
➤ Pronunciation Robustness: Voxtral 3 TTS scores 84.4%, ranking #13 of 32 models, behind Qwen-Audio-3.1-TTS-Plus at 84.6% and ahead of HeyGen Voice at 83.2%. It ranks #3 in Standalone Terms at 93.9%.
➤ Pricing: Voxtral 3 TTS is priced at $13/1M characters, compared to Gemini 3.8 Flash TTS at $16.5/1M characters, Eleven v4 Turbo at $40/1M characters and Simba 3.2 at $6.6/1M characters.
See more details and listen to samples below ⬇️
0
0
0
22
Rafael Aleksandryan retweeted
Stanford researchers did it again.
They made language models 31× faster.
Right now, every major AI, from ChatGPT to Claude, writes exactly the same way. It guesses the next word. One token at a time. Left to right.
It’s like writing an essay on a typewriter while looking computationally exhausting and incredibly slow.
Some labs tried to fix this with "diffusion models”, generating everything at once and polishing it over and over. But that requires massive compute and endless denoising passes.
A team from Stanford just dropped a paper that kills both methods.
It’s called "Next-Concept Modeling" (NCM).
Instead of predicting the next word, the AI predicts the next concept.
It generates text exactly how a professional writer drafts a book. Coarse to fine.
First, it generates a high-level conceptual outline of the entire sequence. Then, in parallel, it fills in the precise words and details across the whole document at once.
The results are staggering.
• Generates text up to 31x faster than current diffusion language models.
• Requires drastically fewer computing steps to reach high-quality outputs.
• Actually outperforms traditional step-by-step models on complex reasoning tasks like Sudoku.
AI is no longer blindly typing word by word, hoping the sentence makes sense when it reaches the period.
It is conceptualizing the entire thought before it opens its mouth.
They made language models 31× faster.
Right now, every major AI, from ChatGPT to Claude, writes exactly the same way. It guesses the next word. One token at a time. Left to right.
It’s like writing an essay on a typewriter while looking computationally exhausting and incredibly slow.
Some labs tried to fix this with "diffusion models”, generating everything at once and polishing it over and over. But that requires massive compute and endless denoising passes.
A team from Stanford just dropped a paper that kills both methods.
It’s called "Next-Concept Modeling" (NCM).
Instead of predicting the next word, the AI predicts the next concept.
It generates text exactly how a professional writer drafts a book. Coarse to fine.
First, it generates a high-level conceptual outline of the entire sequence. Then, in parallel, it fills in the precise words and details across the whole document at once.
The results are staggering.
• Generates text up to 31x faster than current diffusion language models.
• Requires drastically fewer computing steps to reach high-quality outputs.
• Actually outperforms traditional step-by-step models on complex reasoning tasks like Sudoku.
AI is no longer blindly typing word by word, hoping the sentence makes sense when it reaches the period.
It is conceptualizing the entire thought before it opens its mouth.
23
85
465
18,507
Rafael Aleksandryan retweeted
Former $GOOGL employee on why data center capacity and electricity are the real bottlenecks, not GPU production itself, and how unreliable some of the neoclouds are ($AMZN, $GOOGL, $MSFT, $NVDA):
1. According to the expert, open-source models are closing the gap with frontier cted, with customers increasingly replacing Anthropic and OpenAI models in multi-agent architectures with open-source alternatives starting from Q2 of this year. The shift started with simple tasks like tool calling and MCP, but has moved up the stack to low-level orchestration.
2. The expert identifies capacity as the biggest constraint at the moment, with certain companies' traffic growing 5x while GPU availability, particularly in the U.S. where latency requirements are tightest, simply cannot keep up. Sourcing capacity overseas in Asia or Australia is technically possible but adds latency that most customers will not accept.
3. Hyperscalers offer reliability but are expensive, require multi-year commitments of three or more years, and have limited availability of the latest hardware, with customers already chasing B300s. AI-native companies tend to gravitate toward neoclouds due to hyperscaler pricing and contract rigidity, but neoclouds come with their own reliability issues, with GPUs dropping off unexpectedly and requiring engineering resources to manage fleet stability around the clock.
4. The third tier consists of smaller independent operators who have invested in GPU racks, seeing the opportunity to lease out capacity and recoup costs within a year or two, with the remainder being pure profit. The expert describes these as essentially unusable in practice, with operators who do nothing more than find land, install racks, connect electricity, and set up cooling, with no real cloud management, maintenance, or operational expertise behind them.
5. The expert expects overall GPU availability to increase meaningfully over the next year or two, with $NVDA continuing to ramp production. The real bottleneck is not GPUs themselves but data center capacity and electricity in the U.S. On the supply side, the expert also sees players $SPCX beginning to lease out spare capacity to the market, with more likely to follow.
6. The expert notes that U.S.-based neoclouds typically require 20-30% prepayment upfront, with the remainder paid monthly or quarterly over the contract term, and rebates negotiated at the end for any reliability incidents like traffic loss. APAC-based providers operate differently, with no prepayment required at all.
found on @AlphaSenseInc
1. According to the expert, open-source models are closing the gap with frontier cted, with customers increasingly replacing Anthropic and OpenAI models in multi-agent architectures with open-source alternatives starting from Q2 of this year. The shift started with simple tasks like tool calling and MCP, but has moved up the stack to low-level orchestration.
2. The expert identifies capacity as the biggest constraint at the moment, with certain companies' traffic growing 5x while GPU availability, particularly in the U.S. where latency requirements are tightest, simply cannot keep up. Sourcing capacity overseas in Asia or Australia is technically possible but adds latency that most customers will not accept.
3. Hyperscalers offer reliability but are expensive, require multi-year commitments of three or more years, and have limited availability of the latest hardware, with customers already chasing B300s. AI-native companies tend to gravitate toward neoclouds due to hyperscaler pricing and contract rigidity, but neoclouds come with their own reliability issues, with GPUs dropping off unexpectedly and requiring engineering resources to manage fleet stability around the clock.
4. The third tier consists of smaller independent operators who have invested in GPU racks, seeing the opportunity to lease out capacity and recoup costs within a year or two, with the remainder being pure profit. The expert describes these as essentially unusable in practice, with operators who do nothing more than find land, install racks, connect electricity, and set up cooling, with no real cloud management, maintenance, or operational expertise behind them.
5. The expert expects overall GPU availability to increase meaningfully over the next year or two, with $NVDA continuing to ramp production. The real bottleneck is not GPUs themselves but data center capacity and electricity in the U.S. On the supply side, the expert also sees players $SPCX beginning to lease out spare capacity to the market, with more likely to follow.
6. The expert notes that U.S.-based neoclouds typically require 20-30% prepayment upfront, with the remainder paid monthly or quarterly over the contract term, and rebates negotiated at the end for any reliability incidents like traffic loss. APAC-based providers operate differently, with no prepayment required at all.
found on @AlphaSenseInc
104
172
1,126
1,459,165
Implications are huge.
Deepseek is able to push 25-90K tokens/sec per GPU. Far outpacing anyone else.
This is on a single blackwell.
Their margin is insane: $35/hr rev -->$5-10/hr cost.
They make ~70-85% margin at these ludicrously low prices.
They continue to impress.
Deepseek is able to push 25-90K tokens/sec per GPU. Far outpacing anyone else.
This is on a single blackwell.
Their margin is insane: $35/hr rev -->$5-10/hr cost.
They make ~70-85% margin at these ludicrously low prices.
They continue to impress.
🚀 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/De…
1/4
0
0
0
29
This is when @dhh enters his Christian electronic DJ era, coming to a rave near you.
I normally listen to electronic music with few or repeating lyrics when needing to focus, but I've found that a looping Gregorian chant also does the trick! open.spotify.com/track/0xUL1eba…
0
0
0
18
It's interesting....
All of this is like the giving pledge, but actually real and instead of a pledge it's either backed up by facts (money) not the possible generosity of someone who is already dead.
All of this is like the giving pledge, but actually real and instead of a pledge it's either backed up by facts (money) not the possible generosity of someone who is already dead.
Incredible Elon! This will be the largest philanthropic gift in history! 70 M accounts - every citizen under 18 & every child born forevermore now has a free, lifetime investment acct making them an owner in the upside of America & our greatest companies! 🇺🇸🚀🤍@TrumpAccounts @InvestAmerica24
0
0
0
21
Keep in mind that the top OCR accuracy is about 97 on OmniDocBench, so in comparison this is obviously not that great.
That said, amazingly performant use case for say quick document search on your laptop. Bravo to the team., good progress.
That said, amazingly performant use case for say quick document search on your laptop. Bravo to the team., good progress.
Indra OCR 1.1 is out now!
- 1M param main OCR model, 44 pages/s on CPU
- Within 4 points of 3.4B param DeepSeek OCR 2 on OmniParseBench
- 15x faster than Tesseract, and more accurate
- 80+ languages, plus handwriting
1
0
4
250
Impressive.
Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families.
- Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech.
- Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue.
- Ming-3.0-asr: Tiny 8.6B A2B dialects & context.
0
0
0
20
Pretty impressive at bug hunting, but also it's important how things are framed given the model attempted to check a repo that had the actual answers.
Keep in mind Astra also attempted to cheat, just failed.
Keep in mind Astra also attempted to cheat, just failed.
Turns out DeepSeek harness improves the score of DeepSeek V4.1 Flash by 2-3 points on average.
It also costs 80% more due to subagents and more output tokens.
New results:
DeepSeek V4.1 Flash 1017 + Claude Code harness:
- max, mean of 7: 39.1
- high, mean of 5: 36.4
- low, ek V4.1 Flash 1017 + DeepSeek harness:
- max, mean of 7: 42.7
- high, mean of 5: 39.2
- low, mean of 5: 34.8
For comparison:
GPT-6 Astra, max: 45
Opus 5.5, max: 41.7
Muse Wonder Full 1.0, max: 42
It's a crazy model, and the only one catched cheating. All the results are available on the benchmark website.
0
0
0
18