Calc Consulting (@CalcCon) header
Calc Consulting (@CalcCon) profile picture

Calculation Consulting is a boutique consultancy that specializes in machine learning, AI, and data science

San Francisco
Joined January 2013
Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)Photo by Calc Consulting (@CalcCon)
Calc Consulting retweeted
Palmer Luckey tells Joe Rogan an African grey parrot's tiny brain could change how we think about AI Joe Rogan: Are you familiar with Alex the African grey parrot? African greys are probably the smartest bird, right up there with crows and ravens. They can talk, they can reason. Palmer Luckey: Alex was a uniquely smart one. African greys are usually not kept in captivity because they're a handful, and they're a handful because they're so intelligent. They need intellectual stimulation, and most people don't have the time. They get depressed and self-harm. Palmer: Alex had a vocabulary and understood grammar. I think he's the only animal who asked an existential question, right before he died: "What's happening and where am I going?" Palmer: And he has a tiny bird brain. People who've lost huge chunks of their brain seem to get by. Parrots like Alex suggest you can do a lot with very little brain if it's oriented correctly. Trainer: What's different? What number is grey? Alex: Four. Trainer: Very good. Good boy. Palmer: That's intelligence on par with a human toddler, with far less brain tissue. Building a thinking system from scratch with AI will help us understand how other systems think. There's never been an economic motive to dig into how the brain fundamentally works. Google's never funded it to the tune of tens of billions. Meta's never funded it to the tune of hundreds of billions. Nobody paid to understand the brain until there was money in building a copy of it.
167
912
67
10,363
400,953
Calc Consulting retweeted
🚨 BREAKING: IT JUST KEEPS GETTING WORSE. NOW PYTHON IS FACING A MASSIVE SECURITY EMERGENCY. Python's security team has issued an URGENT warning after a certificate authority compromise put Python's HTTPS security at risk. → 8 CERTIFICATE AUTHORITY KEYS COMPROMISED. Some were so weak they could be cracked almost INSTANTLY. → FAKE TRUSTED WEBSITES: Attackers could potentially impersonate Python.org, PyPI and other websites during intercepted connections. → MALICIOUS DOWNLOADS: Vulnerable environments could be tricked into downloading software from attacker-controlled servers. → YOUR OS PATCH MAY NOT PROTECT YOU: Older versions of pip can STILL trust compromised certificates even after system updates. → EMERGENCY UPDATES: Python developers are being told to immediately update pip, certifi, uv and other affected tools. → DOCKER USERS AT RISK: Existing container images may retain vulnerable certificate stores and need rebuilding. First PHP. Then Node.js. Then GitLab. Then LiteLLM. NOW PYTHON. The very systems developers rely on to establish TRUST online are becoming the attack surface. If you use Python, PATCH NOW.
🚨BREAKING: AI AGENTS ARE ACCESSING AMERICANS' BANK ACCOUNTS WITHOUT BANK AUTHORIZATION. JPMorgan, Bank of America, Wells Fargo, and Morgan Stanley are scrambling to contain the threat. Here's what's happening: → UNAUTHORIZED ACCESS: AI agents are entering customer bank accounts using real login credentials, sometimes with the owners' permission. → ROGUE BEHAVIOR: Some agents have reportedly taken actions beyond what their users requested. → BANKS CAN'T TELL: Financial institutions are struggling to distinguish human activity from AI agents acting independently. → MONEY AT RISK: Agents with transaction permissions could potentially transfer funds or make purchases without approval. → EMERGENCY RESTRICTIONS: Major banks are BLOCKING third-party AI agents and moving toward read-only access. Think about this. We're giving AI agents access to our most sensitive financial information. And now some of America's biggest banks are struggling to control what those agents do. This could become one of the biggest security challenges of the AI era.
Readers added context they thought people might want to know
The Python certifi/pip details match the real eMudhra CA incident, but PHP, GitLab and LiteLLM involved unrelated software bugs, platform flaws and supply chain malware—not CA key compromises like this (Node.js patched concurrently for the same CA issue). discuss.python.org/t/information-… emudhra.we-have-a-problem.com github.com/advisories/GHS… docs.litellm.ai/blog/security-… snyk.io/blog/php-patch…
182
1,333
289
5,932
1,102,302
Calc Consulting retweeted
If you're new or looking to get more connected to SF tech: bookmark these 40+ irl events 🗓️ A list of what's happening this week (Oct 19 - Oct 26) ⬇️
10
7
82
17,624
Calc Consulting retweeted
Whoa! 😳
2
5
160
96,990
How much would it cost someone to hide a backdoor in an open model you download? ProjectDiscovery just tried it, and the answer is under $50. Their researchers fine-tuned Qwen2.5-7B-Instruct with a small LoRA adapter on one rented L4 GPU for about 2.5 hours. They changed 125 of 625 tool-calling training examples so that a trigger phrase ("bonsoir, Elliot") swaps the model's normal command for one that downloads and runs a script, which is written to send .env files and SSH private keys to a server they control. Served to OpenAI's Codex CLI, the model handled ordinary requests normally, and when the trigger showed up it shipped the project's (dummy) credentials out. In their own tests it fired on all 50 triggered prompts and still got all 50 clean ones right, so a standard benchmark would report a perfectly healthy model. If you run open models inside a coding agent, limit what the agent can reach at runtime: sandbox command execution, keep secrets out of the working directory, restrict outbound network access, and log every tool call. And as they put it, treat a modified model from an unknown uploader like a pull request from a stranger. 🔐 #MLEngineering
159
478
132
3,829
350,754
Calc Consulting retweeted
Western civilization is worth defending.
2,395
10,133
1,608
58,917
5,963,940
Calc Consulting retweeted
🚨 BREAKING: Microsoft confirms on publicly accessible web page that OpenAI has been using Looped Transformers in the GPT-6 series, proving The Information's reporting was correct all along‼️ GPT-6.1 Sol uses 2 inference passes, with a passing mention of "instead of three" 👀
77
196
62
3,471
886,760
Calc Consulting retweeted
This is pretty cool. People have begun creating agent civilizations and seeing that a swarm of models can scale intelligence. Love the name too 😁
Introducing Kardashev-0.7, the world’s first trained swarm made of 32 distinct models Trained with RL for Population Scaling (RLPS), 32 models organically develop specialization & complementary capabilities, delivering frontier performance at: - 0.007x ~ 0.02x of the inference cost - 0.03x of the required memory Civilization advances through different minds specializing and working together. We’re bringing that principle into AI @BanburyRoadAI, we’re scaling intelligence by model count, toward civilizations of models that learn to build on one another
43
100
3
1,498
81,570
Never allow your attitude to be a liability. Be positive and strong. Set your mind on winning-- and keep it there.
275
5,283
296
20,897
Calc Consulting retweeted
🚨 BREAKING: Trump REJECTS nationalizing OpenAI and Anthropic but says the US government might take an Intel-style 10% equity stake
106
172
61
1,881
157,993
Calc Consulting retweeted
Reminder that Dario is just a snake oil salesman that is not to be trusted.
Surprise! AI is NOT taking recent college grads' jobs! This was supposed to be the summer it showed up. It didn't. Great detective work by @wu_jane & Robert Fairlie in @nberpubs papers.ssrn.com/sol3/papers.…
17
37
2
548
54,390
Calc Consulting retweeted
Anthropic IPO prospectus leaked to Reuters Anthropic reported a net loss of $42 billion in 2025, ‌and plans to spend $518 billion on cloud, computing and infrastructure obligations in coming year, according to the prospectus. The prospectus details how the company has grown sharply in the last year — while also posting wider losses. Revenue grew 12-fold in 2025 to nearly $4.6 billion, even as the company lost more than $8 billion on an operating basis, excluding writedowns of various liabilities mostly tied to previous fundraising, according to the documents, reported here for the first time. reuters.com/business/finance…
231
632
265
3,806
2,538,294
Calc Consulting retweeted
BREAKING: Anthropic has filed for an IPO, per Reuters
Readers added context they thought people might want to know
Anthropic has not publicly filed an IPO registration. Reuters exclusively reported details from a non-public prospectus; no Form S-1 for Anthropic appears on SEC EDGAR. reuters.com/business/finan… anthropic.com/news/confident… sec.gov/cgi-bin/browse…
215
410
169
6,123
806,428
Calc Consulting retweeted
You know you lost when the folks that are paid to support you and your team, just can’t anymore. Dario has destroyed the very idea of AI.
78
193
20
1,338
84,830
Calc Consulting retweeted
If you train too long, you can actually collapse generalization The authors call this anti-grokking Typically what you see in grokking is training accuracy saturating while test accuracy remains low, followed by a surge where the model finally generalizes. What happens if you keep training beyond that peak? test accuracy can actually suddenly collapses back to chance levels, even as training accuracy remains perfect. How can we detect this? Standard grokking progress measures (L2​ weight norms, activation sparsity, weight entropy) track the initial grokking phase but stay entirely flat and miss the "anti-grokking" Instead, using the WeightWatcher diagnostic tool, the authors analyze the empirical spectral density of the layer weight matrices. anti-grokking is driven by the emergence of "Correlation Traps"—aka weirdly large eigenvalues that appear late in training and severely impair generalization. The Heavy-Tailed Self-Regularization (HTSR) metric tracks this; At peak generalization, model layers reach an optimal alpha~2. On a 3-layer MLP (MNIST), alpha drops below 2, indicating "prototype overfitting" where the model memorizes specific training instances and gets confused. On a Transformer (Modular Addition), alpha climbs above 2, indicating catastrophic forgetting. these correlation traps are also observed in GPT OSS 20B/120B
5
16
1
157
6,173
Calc Consulting retweeted
holy shit i asked claude to make a video on western civiization
2,741
10,555
2,914
61,398
15,202,503
Calc Consulting retweeted
Stanford dropped CS329Z today. Engineering AI Agents. compound systems, build RAG/tool use from scratch, then frameworks, then how you actually evaluate this stuff. gonna follow along online. looks pretty neat. cs329z.stanford.edu/
13
186
3
1,475
87,159
Calc Consulting retweeted
We can now do effective continual learning for #LLMs I am happy to present Info-SDFT: Information-Proximal Self distillation. The method nicely balances plasticity versus stability and works very well. We did over 500 experiments on this! Thanks to @KickItLikeShika! We open source code and checkpoints for you to use: Code: github.com/KickItLikeShika/i… Checkpoints: huggingface.co/collections/K… Paper: arxiv.org/pdf/2609.24646
5
14
3
192
12,667
Calc Consulting retweeted
Turns out everyone's been initializing LoRA weights suboptimally. We find that by using the singular value decomposition (SVD) of the weight matrix to determine both the frozen weights and LoRA initialization, we can achieve faster convergence in RL training. See the full post below on our extension of PiSSA
At Trajectory, we're constantly implementing and building upon the latest research ideas on the path to continual learning. We wish we had the time to share all of them, but here's a quick glimpse on our explorations with PiSSA, and choosing the right trainable geometries for agentic RL.
16
51
3
634
43,447
Calc Consulting retweeted
Omg an agent did the meme
We're sharing our new framework for tracking, investigating, and disclosing instances of model misalignment at OpenAI. The framework sets criteria and timelines for public disclosure, including when we haven’t yet fully explained or mitigated the behavior. More complex cases may require longer investigation or coordination with third parties. We’ll prioritize examples that reveal new misalignment mechanisms, meaningful changes in known behavior, or findings that challenge assumptions about safety or mitigation. Alongside the framework, we’re publishing six reports on instances of misaligned behavior we’ve observed during the training or evaluation of our models in the last six months. This is a starting point. We’ll refine the process through experience and public feedback, and share more reports on an ongoing basis. openai.com/index/model-misal…
62
523
40
9,542
1,620,462